{ "_comment": "One brief gpt-image-2 image-edit prompt per task. Prompts re-render the v4_plain matplotlib source as a photorealistic / natural variant while preserving the task's discriminative structure (positions, counts, labels). One sample (#00000) generated per task for review.", "tasks": { "distributed_scanning/attribute_group_counting": { "src": "images/attribute_group_counting_00000.png", "prompts": [ "Re-render as a photorealistic top-down photograph of stones laid on a NOISY natural surface — pick ONE: damp riverbed gravel, mossy forest floor with twigs and leaves, dark wet beach sand with shell fragments, weathered wooden plank with knots and scuffs.\n\nEach blob in the source becomes ONE stone with the EXACT same silhouette, size, and position. ALL stones share the SAME shading, SAME colour, and SAME surface texture (one stone material across the whole image — e.g. uniformly dark slate, or uniformly pale river pebble). Stones differ ONLY in OUTLINE SHAPE; nothing else distinguishes them. Subtle natural lighting micro-variation per stone is fine, but no recolouring, no different materials, no patterned tops.\n\nThe stones must contrast STRONGLY with the surface (uniform dark stones on a pale sand surface, or uniform pale stones on dark moss) so each silhouette outline is unambiguously visible to a human at a glance. Background clutter (loose grain, dust, twigs) is allowed but must NEVER produce another stone-shaped blob and must NEVER intrude on the stone outlines.\n\nDo not add or remove any stone. No text, no labels, no overlays.", "Re-render as a photorealistic overhead photograph of cookies laid out on a NOISY baker's surface — pick ONE: flour-dusted dark wooden board with crumbs, parchment-lined baking tray with butter smears, scratched dark stone counter with chocolate flakes, worn metal sheet with sugar specks.\n\nEach blob in the source becomes ONE cookie with the EXACT same silhouette, size, and position. ALL cookies share the SAME shading, SAME dough colour, and SAME surface texture (one cookie type across the whole image — e.g. uniformly golden shortbread, or uniformly dark chocolate cookie). Cookies differ ONLY in OUTLINE SHAPE; nothing else distinguishes them. Subtle bake-variation in tone is fine, but no different cookie types, no contrast garnishes inside.\n\nThe cookies must contrast STRONGLY with the surface (golden cookies on a dark counter, or dark cookies on a pale parchment) so every silhouette outline is unambiguously visible. Background clutter (crumbs, flour patches, butter smears) is allowed but must NEVER produce another cookie-shaped blob and must NEVER intrude on the cookie outlines.\n\nDo not add or remove any cookie. No text, no labels, no overlays.", "Re-render as a photorealistic close-up overhead photograph of leaves scattered on a NOISY forest floor — pick ONE: damp dark soil with twigs and pine needles, mossy bark with lichen patches, dry oak-leaf litter, fallen autumn leaves with bark fragments.\n\nEach blob in the source becomes ONE leaf with the EXACT same silhouette, size, and position. ALL leaves share the SAME shading, SAME colour, and SAME surface texture (one leaf species across the whole image — e.g. uniformly bright yellow ginkgo, or uniformly dark green ivy). Leaves differ ONLY in OUTLINE SHAPE; nothing else distinguishes them. Subtle natural weathering per leaf is fine, but no different species, no veining variation that creates internal categories.\n\nThe leaves must contrast STRONGLY with the surface (bright yellow leaves on dark damp soil, or dark green leaves on pale dry mulch) so each silhouette outline is unambiguously visible. Background natural clutter (twigs, needles, dirt, lichen) is allowed but must NEVER produce another leaf-shaped blob and must NEVER intrude on the leaf outlines.\n\nDo not add or remove any leaf. No text, no labels, no overlays.", "Re-render as a photorealistic top-down photograph of porcelain tea cups laid on a NOISY domestic surface — pick ONE: a worn wooden table with crumbs and a faint coffee ring, a marble counter with sugar specks, a chipped enamel tray with droplets, an old slate counter with smudges and dust.\n\nEach blob in the source becomes ONE porcelain tea cup with the EXACT same silhouette, size, and position. ALL tea cups share the SAME shading, SAME glaze colour (e.g. uniformly white-bone porcelain, OR uniformly pale-blue glaze), and SAME surface texture across the whole image. Cups differ ONLY in their OUTLINE SHAPE (the silhouette of the rim seen from above varies — round, oval, slightly squared); nothing else distinguishes them. Subtle natural lighting micro-variation per cup is fine, but no painted designs, no different glazes, no decorative rims that create internal categories.\n\nThe cups must contrast STRONGLY with the surface (white porcelain on dark wood, or dark glaze on pale marble) so each silhouette outline is unambiguously visible to a human at a glance. Background clutter (crumbs, dust, smudges) is allowed but must NEVER produce another cup-shaped blob and must NEVER intrude on the cup outlines.\n\nDo not add or remove any cup. No text, no labels, no overlays.", "Re-render as a photorealistic top-down photograph of hand-pressed clay tiles laid on a NOISY workshop surface — pick ONE: a dusty potter's wooden bench with stray clay scraps, a stone-flagged kiln floor with ash flecks, a worn metal tray with terracotta dust, a chipped concrete slab with tool marks.\n\nEach blob in the source becomes ONE hand-pressed clay tile with the EXACT same silhouette, size, and position. ALL tiles share the SAME shading, SAME terracotta colour, and SAME surface texture (one clay material across the whole image — uniformly warm terracotta, slightly matte). Tiles differ ONLY in OUTLINE SHAPE; nothing else distinguishes them. Subtle firing variation in tone is fine, but no glazes, no painted patterns, no inlays that create internal categories.\n\nThe tiles must contrast STRONGLY with the surface (terracotta on dark stone, or dark tiles on pale concrete) so each silhouette outline is unambiguously visible to a human at a glance. Background clutter (clay scraps, dust, tool marks) is allowed but must NEVER produce another tile-shaped blob and must NEVER intrude on the tile outlines.\n\nDo not add or remove any tile. No text, no labels, no overlays." ] }, "distributed_scanning/bounded_faces_counting": { "src": "images/bounded_faces_counting_00000.png", "prompts": [ "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE driftwood stake hammered into a salt flat at dawn at the EXACT same position. Each line segment becomes ONE old rope strung taut between the SAME pair of stakes, following the line exactly with no slack, no extras, no missing strands. NOISY surface: cracked salt crusts, wet rust streaks, scattered shell fragments, drift-debris, faint footprints. Stakes and ropes sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE rusted iron rebar peg driven into dry cracked earth at the EXACT same position. Each line segment becomes ONE weathered hemp twine knotted taut between the SAME pair of pegs, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry red dust, scattered pebbles, clumps of dead grass, tyre tracks, paint flecks, occasional metal shavings. Rebar and twine sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE mossy stone bollard on a damp riverbank at the EXACT same position. Each line segment becomes ONE braided seaweed strand stretched taut between the SAME pair of bollards, following the line exactly with no slack, no extras, no missing strands. NOISY surface: damp pebbles, fallen leaves, lichen crust, twigs, water-pooled grit, micro-puddles. Stones and seaweed sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE wooden tent stake hammered into a dry meadow at the EXACT same position. Each line segment becomes ONE taut nylon guy-line strung between the SAME pair of stakes, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry yellow grass tufts, scattered wildflowers, twigs, faint footpaths, occasional pebbles. Stakes and guy-lines sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE weathered wooden telegraph pole stub seen from directly above standing in cracked earth at the EXACT same position. Each line segment becomes ONE sagging black copper wire stretched between the SAME pair of pole tops, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry red soil with deep cracks, scattered scrub-grass tufts, pebbles, tyre tracks, rust flecks. Pole stubs and wires sit visibly on top, never hidden by clutter. No text, no labels, no overlays." ] }, "distributed_scanning/counting_connected_components": { "src": "images/counting_connected_components_00000.png", "prompts": [ "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE driftwood stake hammered into a salt flat at dawn at the EXACT same position. Each line segment becomes ONE old rope strung taut between the SAME pair of stakes, following the line exactly with no slack, no extras, no missing strands. NOISY surface: cracked salt crusts, wet rust streaks, scattered shell fragments, drift-debris, faint footprints. Stakes and ropes sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE rusted iron rebar peg driven into dry cracked earth at the EXACT same position. Each line segment becomes ONE weathered hemp twine knotted taut between the SAME pair of pegs, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry red dust, scattered pebbles, clumps of dead grass, tyre tracks, paint flecks, occasional metal shavings. Rebar and twine sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE mossy stone bollard on a damp riverbank at the EXACT same position. Each line segment becomes ONE braided seaweed strand stretched taut between the SAME pair of bollards, following the line exactly with no slack, no extras, no missing strands. NOISY surface: damp pebbles, fallen leaves, lichen crust, twigs, water-pooled grit, micro-puddles. Stones and seaweed sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE wooden tent stake hammered into a dry meadow at the EXACT same position. Each line segment becomes ONE taut nylon guy-line strung between the SAME pair of stakes, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry yellow grass tufts, scattered wildflowers, twigs, faint footpaths, occasional pebbles. Stakes and guy-lines sit visibly on top, never hidden by clutter. No text, no labels, no overlays.", "Re-render as a wild, weathered, real-world top-down photograph. Each dot in the source becomes ONE weathered wooden telegraph pole stub seen from directly above standing in cracked earth at the EXACT same position. Each line segment becomes ONE sagging black copper wire stretched between the SAME pair of pole tops, following the line exactly with no slack, no extras, no missing strands. NOISY surface: dry red soil with deep cracks, scattered scrub-grass tufts, pebbles, tyre tracks, rust flecks. Pole stubs and wires sit visibly on top, never hidden by clutter. No text, no labels, no overlays." ] }, "distributed_scanning/counting_regions": { "src": "images/counting_regions_00000.png", "prompts": [ "Re-render as a photorealistic top-down aerial photograph of a vast UNIFORM grass meadow, where every region shares the EXACT same tall green grass texture. Inside every region the ground is identical mature meadow grass with no biome variation, no colour variation, no density variation, no patches of any kind.\n\nBoundaries are the ONLY feature separating regions, and they are drawn as ANIMAL TRAILS — thin worn paths threaded through the grass where countless small animals (deer, foxes, badgers) have walked over time. Each trail follows the source partition curve EXACTLY. The trails are deliberately HARD for naive computer-vision tools to detect:\n • The grass on BOTH sides of a trail is the same height, same colour, same species — only the trail itself shows slightly trampled grass and bare soil glimpses.\n • Trail width is irregular along its length — sometimes 4 cm, sometimes 8 cm, occasionally pinched almost shut where grass has grown back.\n • Trails are NOT smooth painted lines: they are noisy, broken, partly overgrown, with stray grass blades crossing them.\n • There are NO hard high-contrast edges anywhere: the soil shows through only in faint patches.\n • The grass texture must remain dominant across the entire image; the trails are subtle, narrow, and nearly camouflaged — a careful human can trace them by following the slightly trampled line, but a thresholding/edge-detection algorithm would see only one big texture.\n\nTrails must trace the source partition curve faithfully on every twist and turn, with NO gaps, NO breaks, NO doubled lines, NO new trails inside a region. Adjacent regions are distinguishable ONLY by spotting these subtle trails between them.\n\nNo text, no labels, no compass, no overlays, no animals visible, no people, no fences, no other features.", "Re-render as a photorealistic top-down aerial photograph of a UNIFORM expanse of fresh white snow at dawn, where every region shares the EXACT same fresh-snow texture. Inside every region the surface is identical undisturbed powdery snow — no biome variation, no colour variation, no patches of any kind.\n\nBoundaries are the ONLY feature separating regions, and they are drawn as SINGLE-FILE FOOTPRINT TRAILS — a line of HIGHLY IRREGULAR footprints (mix of adult boots of different sizes, smaller children's boots, bare-foot prints, animal paw-prints from foxes/deer/rabbits, snowshoe imprints) where various walkers crossed along each boundary. Each trail follows the source partition curve EXACTLY. The trails are deliberately HARD for naive computer-vision tools to detect:\n • The snow on BOTH sides of a footprint trail is the same fresh untouched powder — only the prints themselves show as small dimpled depressions in the snow with soft blue-grey shadows.\n • The footprints are DISCRETE — small rounded boot shapes spaced ~25–40 cm apart along the trail, NOT a continuous line. To follow a trail a viewer must mentally CONNECT THE DOTS along the source curve.\n • Footprint SHAPE varies along each trail (different boots, different paws, even occasional bare-foot or snowshoe), and depth, orientation, and stride length vary too. The trail meanders slightly off the source curve (within ~3 cm) before snapping back (some prints sharper, some softer where wind has dusted snow back into them).\n • There are NO hard linear edges anywhere; each footprint is an isolated small feature, only their collective alignment along the source curve reveals the boundary.\n • The snow surface ALSO contains MANY NOISY DISTRACTORS unrelated to any boundary: scatter ~80–150 random isolated features across the entire canvas — stray paw-prints in random positions, dropped twigs, fallen pine cones, small icy lumps, dimpled wind hollows, scattered red berries, fallen brown leaves blown over the snow, occasional small rabbit-droppings. These distractors are randomly placed and NOT aligned with any partition curve. The vast majority of the image is still uniform fresh snow — a careful human can trace each boundary by visually chaining its footprints, but a basic edge-detector would just register thousands of unrelated little blobs.\n\nFootprint trails must trace the source partition curve faithfully on every twist and turn, with NO gaps too long for a human to chain (max ~50 cm between successive prints), NO doubled trails, NO new trails inside a region. Adjacent regions are distinguishable ONLY by spotting these subtle dotted lines.\n\nNo text, no labels, no compass, no overlays, no people visible, no animals, no other features.", "Re-render as a photorealistic top-down close-up photograph of one large continuous patch of dry beige sand on a beach, covering the ENTIRE canvas. The sand has natural organic variation: subtle ripples from wind, scattered grain-size differences, micro-shadow patterns, occasional tiny pebbles, broken shell fragments, mica flakes, dropped sesame-sized debris, beetle exoskeletons, dried seaweed bits, salt crystals — strew ~150–250 small dark distractor specks randomly across the canvas, NOT aligned with any partition curve — these texture features flow continuously across the whole image with NO RESPECT for the partition.\n\nThe texture variation is INDEPENDENT of the regions: a wind-ripple band can sit anywhere on the sand, span across multiple regions, or fall entirely inside one region. Inside every region the sand has the EXACT same character (same beige hue, same grain texture, same ambient ripple distribution).\n\nBoundaries are the ONLY feature separating regions, drawn as TINY ANT MARCH LINES — a single-file column of small black ants moving along the boundary. Each line follows the source partition curve EXACTLY. The trails are deliberately HARD for naive computer-vision tools to detect:\n • The sand on BOTH sides of every ant march is identical — only the ants themselves break the uniform sand texture.\n • The ants are DISCRETE small black specks (1–2 mm visible diameter from above) spaced 2–4 mm apart along the line. They do NOT form a continuous edge — to read a boundary the viewer must mentally CHAIN successive ants along the source curve.\n • Some ants are crisp, some blurry from motion; gaps occasionally widen up to 1 cm where the column thins; occasional outlier ants stray briefly off the line and back.\n • The image must look CHAOTICALLY NOISY — hundreds of small dark dots scattered everywhere, only a small fraction of them being ants on the partition curves. There is NO continuous high-contrast edge anywhere — only thousands of small dark specks. A basic edge-detector returns just blob noise without alignment reasoning; a careful human reads the line by visual gestalt.\n\nAnt march lines must trace the source partition curve faithfully on every twist and turn, with NO gaps too long for a human to chain (max ~1 cm between successive ants), NO doubled lines, NO new lines inside a region. Adjacent regions are distinguishable ONLY by chaining these dotted ant trails.\n\nNo text, no labels, no compass, no overlays, no other features.", "Re-render as a photorealistic top-down macro photograph of a single sheet of aged cream-coloured parchment paper covering the ENTIRE canvas. The parchment surface has natural variation across the whole image — subtle tea-staining, faint fold creases, water spots, and aged fibres flowing continuously, with NO RESPECT for the partition.\n\nCRITICAL — the distractors and the boundary are the SAME material (dark brown/black ink) but DIFFERENT SHAPE CATEGORIES:\n • Scatter 30–50 SHORT INK MARKS across the parchment, randomly positioned and randomly oriented: small ink blots, brief signatures, single-letter scribbles, smudges, fingerprint smears, short underline strokes (each 0.5–4 cm long). They look like accidental marks left by hand. They are SHORT and DO NOT trace the partition curve.\n • Add 10–20 SLIGHTLY LONGER ink doodles — small loops, simple flourishes, abstract squiggles (each 4–8 cm). Still localised; do NOT span across the canvas.\n • All marks are SAME ink colour and SAME ink texture as the boundary (so a shape-classifier sees them as the same object class).\n\nBoundaries must trace the source partition curve, but the BOUNDARY ITSELF MUST BE NOISY (not a crisp continuous line) so that classical edge / Hough / line-detection CV pipelines fail:\n • The boundary is rendered as a SERIES OF DISJOINT INK DASHES of irregular length (each 1–3 cm), with small visible GAPS between dashes (0.3–1 cm gaps), so the line is BROKEN as a pixel-string.\n • Dash thickness VARIES along the curve (1–4 mm), and ink OPACITY varies (some dashes dark and saturated, others faint and washed-out as if the quill was running dry).\n • Occasional ink BLOTS, small smudges, and tiny accidental flicks deviate slightly off the partition curve (within 1–2 mm of the true curve), so the boundary has a hand-drawn, noisy feel rather than a clean drafted line.\n • Despite the noise, every dash sits ON the source partition curve so a human can still TRACE the curve by chaining the dashes visually. Wherever a partition curve touches the canvas border in the source, the dashed boundary in the output must EXTEND ALL THE WAY TO THE CORRESPONDING IMAGE EDGE — do NOT inset, crop, fade out, or stop the dashes before the edge. Dashes continue right up to the page edge.\n • The boundary is the SAME ink colour and texture as the distractor scribbles, distinguishable from them ONLY by the dashes' COLLECTIVE alignment along the partition curve through every twist and turn.\n\nThe boundary must follow the source partition faithfully on every twist with NO doubled lines and NO new boundaries inside any region. Distractor scribbles must NOT trace the partition.\n\nA naive edge-detector finds a noisy mess of unconnected ink fragments. A careful human reads the boundary by gestalt continuity along each curve.\n\nNo text other than the ink scribbles, no labels, no compass, no overlays, no other features.", "Re-render as a photorealistic top-down close-up photograph of a wet sandy beach at low tide. The sand — pale beige, faint wind ripples, micro shadows, continuous grain texture — fills the ENTIRE canvas, flowing continuously with NO RESPECT for the partition.\n\nCRITICAL — the beach is strewn with many ROPE-LIKE LINEAR distractors that look visually similar to the real boundary rope:\n • Scatter 30–60 discarded SHORT ROPE SCRAPS of varied lengths (10 cm – 1.5 m), varied colours and weaves (manila / sisal / nylon / faded synthetic), randomly oriented across the canvas. They are short and DO NOT trace the partition curve.\n • Scatter 30–60 long pieces of SEAWEED, KELP STRANDS, DRIFTWOOD, AND DRIED VINES — all narrow linear features, looking ropy at first glance, randomly oriented across the canvas, NOT aligned with the partition.\n • Add 20–40 ropy crab-track grooves and worm-trail marks in the sand — narrow squiggly lines randomly placed, NOT aligned with the partition.\n • Add 60–120 scattered shells, pebbles, and small drift debris for general clutter.\n\nSo the image contains DOZENS of rope-like and linear-strand-like features, none of which — except one — trace the partition. Edge-detection / line-detection / shape-classifier CV will find HUNDREDS of false positives.\n\nBoundaries are the ONLY feature that traces the partition: a CHAIN OF MEDIUM-LENGTH MANILA ROPE SEGMENTS (each 25–60 cm long, ~2 cm thick, weathered golden-brown, twisted-fibre texture) lying flat on the sand, end-to-end with VERY SMALL gaps (5–15 mm) between successive segments. The boundary reads as a CLEARLY CONTINUOUS rope trail to a human — each segment touches or nearly touches the next, with only a hairline gap. The gaps are small enough that a human eye instantly chains them as one rope, while still being technically broken to defeat continuous-line CV detectors. Together the segments trace the source partition curve EXACTLY through every twist and turn. The boundary rope segments are visually nearly identical to the many short rope-scrap distractors (same material, same thickness, same texture) — distinguishable from them ONLY by their COLLECTIVE alignment along the partition curve.\n\nA human can trace the boundary easily because the rope segments are long and barely separated, reading as a near-continuous rope line. A naive line/shape detector cannot, because dozens of similar-looking ropy features lie everywhere.\n\nThe disjoint rope segments must collectively trace the source partition curve faithfully on every twist; gaps between segments are intentional but very small (5–15 mm) so the rope reads as continuous to humans. NO doubled lines. There must be NO new boundary ropes inside any region. The rope must NEVER be hidden under clutter — it sits clearly visible on top.\n\nNo text, no labels, no compass, no overlays, no people, no boats." ] }, "distributed_scanning/tangled_loops": { "src": "images/tangled_loops_00000.png", "prompts": [ "Re-render as a photorealistic top-down photograph of ropes lying tangled on a NOISY real-world surface — pick ONE: an old workshop floor with scattered sawdust, screws, and grease patches; a damp forest floor with fallen leaves, twigs, moss, and small pebbles; a worn warehouse concrete floor stained with paint splatters, grime, and dust; a low-tide beach with shells, kelp scraps, and wet sand ripples. The surface should look genuinely textured and lived-in (real-world camera grain noise, varied lighting), not a clean studio backdrop.\n\nALL loops must be the SAME uniform colour and SAME material (one type of rope) chosen to CONTRAST sharply with the surface so every strand stays unambiguously readable on top. Each loop follows the EXACT same path as in the source, with proper over/under crossings at every intersection. Loops must NEVER be hidden, broken, or interrupted by background clutter — they sit visibly on top.\n\nPreserve every loop. Do not add or remove any loop. No text, no labels, no overlays.", "Re-render as a photorealistic top-down photograph of tangled cords lying on a NOISY cluttered real-world surface — pick ONE: a workshop floor strewn with sawdust, screws, and metal shavings; a damp forest floor with fallen leaves, twigs, and moss; an attic floor with dust, old papers, and crumpled boxes; a beach at low tide with shells, kelp, and pebbles. ALL loops must be the SAME uniform colour and SAME material (e.g., bright orange paracord, glossy black rubber cable, white nylon rope) chosen to CONTRAST sharply with the noisy background so each strand remains unambiguously readable on top. Each loop follows the EXACT same path as in the source, with proper over/under crossings at every intersection. Cords must NEVER be hidden or interrupted by background clutter. Preserve every loop. Do not add or remove any loop. No text, no labels.", "Re-render as a photorealistic top-down photograph of a CHAOTIC pile of tangled identical objects on a busy real-world setting — pick ONE: bright orange extension cords on a workshop bench scattered with tools, gaffer tape, and dust; identical silver chains spilled across a black velvet jewellery tray cluttered with mixed beads, buttons, and clasps; coiled garden hoses on a patio strewn with leaves, sprinkler heads, and gravel; black stage cables on a backstage floor with crumpled tape, paper, and shoe scuffs. ALL loops must be the SAME uniform colour and SAME material; the surrounding clutter contains many similar-looking but DIFFERENT short bits (frayed ends, smaller cords, scrap pieces) that are NOT part of any loop. Each real loop follows the EXACT same path as in the source, with proper over/under crossings; the loops always sit visibly on top of the clutter and are never broken by it. Preserve every loop. Do not add or remove any loop. No text, no labels.", "Re-render as a photorealistic top-down photograph of bright climbing rope tangled on a NOISY rocky alpine ledge — scattered carabiners, chalk dust, small pebbles, lichen patches, snow-melt streaks. ALL loops must be SAME uniform colour and material (e.g. high-vis orange dynamic climbing rope, OR royal-blue accessory cord — pick ONE per image, same across every loop) chosen to CONTRAST sharply with the rock and clutter so each strand stays unambiguously readable on top. Each loop follows the EXACT same path as in the source, with proper over/under crossings at every intersection. Loops must NEVER be hidden, broken, or interrupted by background clutter — they sit visibly on top.\n\nPreserve every loop. Do not add or remove any loop. No text, no labels, no overlays.", "Re-render as a photorealistic top-down photograph of dark electrical cables piled on a NOISY backstage floor — strewn gaffer-tape scraps, set-piece edges, tape gel fragments, scuff marks, leftover rosin dust. ALL loops must be SAME uniform colour and material (e.g. matte-black rubber-jacketed XLR cable, OR thick yellow stage extension cord — pick ONE per image, same across every loop) chosen to CONTRAST sharply with the floor and clutter so each strand stays unambiguously readable on top. Each loop follows the EXACT same path as in the source, with proper over/under crossings at every intersection. Cables must NEVER be hidden, broken, or interrupted by background clutter — they sit visibly on top.\n\nPreserve every loop. Do not add or remove any loop. No text, no labels, no overlays." ] }, "sequential_traversal/arrow_chain": { "src": "images/arrow_chain_00000.png", "prompts": [ "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • Every footprint stays at the EXACT same (x, y) within ~2 px AND points in the EXACT same direction within ~3° — do NOT move, rotate, mirror, scale, or re-orient any footprint.\n • Every labeled circle keeps the SAME letter and the SAME (x, y) position. Letter glyphs unchanged.\n • Object COUNT equals the source EXACTLY — never add, skip, or merge.\n • Position and direction are the load-bearing signal of this image; nothing about geometry may change.\n • The single chain cell with a GREEN ring around it is the START marker. Preserve the green ring CLEARLY in the rendered image — a bright vivid green circle of the SAME size and at the SAME position as in the source. Do NOT remove, recolour, or fade this green start ring; it is critical to the question.\n\nRe-render as a photorealistic top-down photograph of damp beach sand. Every footprint in the source becomes ONE realistic shoe/boot footprint pressed into the sand, in the SAME position, pointing the SAME direction (toes = direction). Each labeled circle becomes a flat, smooth wet pebble with the SAME letter painted in white. Sand has natural ripples and grain texture; lighting is even and soft. No text other than the letters on pebbles.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • Every fish stays at the EXACT same (x, y) within ~2 px AND points in the EXACT same direction within ~3° — do NOT move, rotate, mirror, scale, or re-orient any fish.\n • Every labeled circle keeps the SAME letter and the SAME (x, y) position. Letter glyphs unchanged.\n • Object COUNT equals the source EXACTLY — never add, skip, or merge.\n • Position and direction are the load-bearing signal of this image; nothing about geometry may change.\n • The single chain cell with a GREEN ring around it is the START marker. Preserve the green ring CLEARLY in the rendered image — a bright vivid green circle of the SAME size and at the SAME position as in the source. Do NOT remove, recolour, or fade this green start ring; it is critical to the question.\n\nRe-render as a photorealistic top-down photograph of clear shallow tropical water over white sand. Every fish in the source becomes ONE small tropical fish swimming flat over the sand at the SAME position, pointing the SAME direction (head = direction). Each labeled circle becomes a flat round coral disc bearing the SAME letter. Water shows soft caustic ripples; sand has natural grain. No text other than the letters on coral.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • Every bird stays at the EXACT same (x, y) within ~2 px AND points in the EXACT same direction within ~3° — do NOT move, rotate, mirror, scale, or re-orient any bird.\n • Every labeled circle keeps the SAME letter and the SAME (x, y) position. Letter glyphs unchanged.\n • Object COUNT equals the source EXACTLY — never add, skip, or merge.\n • Position and direction are the load-bearing signal of this image; nothing about geometry may change.\n • The single chain cell with a GREEN ring around it is the START marker. Preserve the green ring CLEARLY in the rendered image — a bright vivid green circle of the SAME size and at the SAME position as in the source. Do NOT remove, recolour, or fade this green start ring; it is critical to the question.\n\nRe-render as a photorealistic top-down photograph of a flock of small birds flying over a pale overcast sky. Every bird in the source becomes ONE small bird in flight at the SAME position, beak/head pointing the SAME direction. Each labeled circle becomes a small painted wooden disc bearing the SAME letter, suspended in the same plane. Sky has subtle cloud variation; lighting is even. No text other than the letters on discs.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • Every key stays at the EXACT same (x, y) within ~2 px AND points in the EXACT same direction within ~3° — do NOT move, rotate, mirror, scale, or re-orient any key.\n • Every labeled circle keeps the SAME letter and the SAME (x, y) position. Letter glyphs unchanged.\n • Object COUNT equals the source EXACTLY — never add, skip, or merge.\n • Position and direction are the load-bearing signal of this image; nothing about geometry may change.\n • The single chain cell with a GREEN ring around it is the START marker. Preserve the green ring CLEARLY in the rendered image — a bright vivid green circle of the SAME size and at the SAME position as in the source. Do NOT remove, recolour, or fade this green start ring; it is critical to the question.\n\nRe-render as a photorealistic top-down photograph of an old wooden workbench. Every key in the source becomes ONE antique brass key lying flat on the wood at the SAME position, with the bow end pointing the SAME direction (the pointy/shaft end). Each labeled circle becomes a small leather tag stamped with the SAME letter. Wood has natural grain and minor scratches; lighting is warm and soft. No text other than the letters on tags.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • Every leaf stays at the EXACT same (x, y) within ~2 px AND points in the EXACT same direction within ~3° — do NOT move, rotate, mirror, scale, or re-orient any leaf.\n • Every labeled circle keeps the SAME letter and the SAME (x, y) position. Letter glyphs unchanged.\n • Object COUNT equals the source EXACTLY — never add, skip, or merge.\n • Position and direction are the load-bearing signal of this image; nothing about geometry may change.\n • The single chain cell with a GREEN ring around it is the START marker. Preserve the green ring CLEARLY in the rendered image — a bright vivid green circle of the SAME size and at the SAME position as in the source. Do NOT remove, recolour, or fade this green start ring; it is critical to the question.\n\nRe-render as a photorealistic top-down photograph of forest floor with damp earth and moss. Every leaf in the source becomes ONE real fallen leaf lying flat on the soil at the SAME position, with its TIP pointing the SAME direction as in the source (stem opposite the tip). Each labeled circle becomes a small smooth river stone bearing the SAME letter painted in white. The soil has scattered fine debris (twig fragments, soil grains); lighting is overcast and even. No text other than the letters on stones." ] }, "sequential_traversal/color_zone_sequence": { "src": "images/color_zone_00000.png", "prompts": [ "Re-render as a hand-drawn stylised tourist map (think: fantasy / RPG world map, or vintage national-park trail brochure). The output should look LIKE A DRAWN MAP — flat colour regions, clean borders, decorative shading — not a photorealistic aerial photo. Stylisation makes precise geometry easier to preserve.\n\nGEOMETRY MUST BE PRESERVED EXACTLY:\n • Each coloured region stays at the EXACT same shape, position, and adjacency. Replace each region's flat colour with one biome icon-fill (forest dotted with little trees / lake with light wave hatching / meadow with grass tufts / desert with sand dots / mountains with peak symbols). Choose ONE biome per region; do NOT mix biomes inside a region; do NOT bleed a biome across a boundary.\n • Boundaries between regions are thin dark hand-drawn ink lines, traced EXACTLY along the source partition curves.\n • Every black curve in the source becomes a SINGLE UNIFORM-STYLE trail in the output. ALL curves — the query curve AND every distractor curve — must be rendered IDENTICALLY: same colour, same dash pattern, same line weight, same texture. The viewer must NOT be able to tell which curve is the query just by looking at the lines themselves; the only distinguishing feature allowed is the small black starting dot on the query curve. Trace each source curve EXACTLY (do NOT smooth, shorten, reroute, or break any curve). • Each single-letter region label stays at the EXACT same position with the SAME glyph, drawn in a small map-label box.\n\nNo new regions, no new trails, no new labels. No photorealism, no 3D perspective. Pure flat illustrated map look. Do not colour-code or style-code the trails differently from each other.\n\nThe source image contains a literal 'Start' text label inside a small white box with a thin black border, placed at the beginning of the query curve. PRESERVE THIS 'Start' LABEL EXACTLY AS-IS in the output: same wording (\"Start\"), same position, same small white box with thin black border, same legibility. Do NOT replace the label with any icon, do NOT translate the word, do NOT restyle the box. The 'Start' label is the ONLY distinguishing mark between the query curve and the distractor curves; every curve must otherwise look identical (same color, same dash pattern, same weight, same texture).", "Re-render as a vintage maritime nautical chart on aged parchment (think: 18th-century sailor's chart with calligraphic flourishes). Flat-style cartography — NOT a photorealistic ocean photo. Stylisation makes precise geometry easier to preserve.\n\nGEOMETRY MUST BE PRESERVED EXACTLY:\n • Each coloured region stays at the EXACT same shape, position, and adjacency. Replace each region's flat colour with ONE distinctive island-or-sea texture: deep ocean with tiny waves / sandbar with fine stippling / coral reef with cross-hatching / mangrove with leaf flecks / kelp forest with curling strokes / open beach with dotted sand. ONE texture per region; do NOT mix textures inside a region; do NOT bleed a texture across a boundary.\n • Boundaries between regions are thin sepia-ink hand-drawn coastlines, traced EXACTLY along the source partition curves.\n • Every black curve in the source becomes a SINGLE UNIFORM-STYLE rhumb line in the output. ALL curves — the query curve AND every distractor curve — must be rendered IDENTICALLY: same sepia ink colour, same uniform thin solid stroke, same weight, same texture. The viewer must NOT be able to tell which curve is the query just by looking at the lines themselves; the only distinguishing feature allowed is the literal text 'Start' label sitting on the query curve. Trace each source curve EXACTLY (do NOT smooth, shorten, reroute, or break any curve). • Each single-letter region label stays at the EXACT same position with the SAME glyph, drawn in a small calligraphic cartouche.\n\nNo new regions, no new lines, no new labels. No photorealism, no 3D perspective. Pure flat antique chart look. Do not colour-code or style-code the rhumb lines differently from each other.\n\nThe source image contains a literal 'Start' text label inside a small white box with a thin black border, placed at the beginning of the query curve. PRESERVE THIS 'Start' LABEL EXACTLY AS-IS in the output: same wording (\"Start\"), same position, same small white box with thin black border, same legibility. Do NOT replace the label with any icon, do NOT translate the word, do NOT restyle the box. The 'Start' label is the ONLY distinguishing mark between the query curve and the distractor curves; every curve must otherwise look identical (same color, same dash pattern, same weight, same texture).", "Re-render as a flat children's-storybook watercolour map (think: a friendly hand-painted nursery storybook). Flat illustrated style — NOT a photorealistic aerial photo. Stylisation makes precise geometry easier to preserve.\n\nGEOMETRY MUST BE PRESERVED EXACTLY:\n • Each coloured region stays at the EXACT same shape, position, and adjacency. Replace each region's flat colour with ONE storybook biome wash (a meadow of tiny flowers / a cosy forest of round trees / a candy-pink desert with cacti / a pastel pond with lily pads / a soft mountain pass with tiny rocks / a snow patch with little snowflakes). ONE biome per region; do NOT mix biomes inside a region; do NOT bleed a biome across a boundary.\n • Boundaries between regions are thin charcoal hand-drawn outlines, traced EXACTLY along the source partition curves.\n • Every black curve in the source becomes a SINGLE UNIFORM-STYLE pencil dotted path in the output. ALL curves — the query curve AND every distractor curve — must be rendered IDENTICALLY: same charcoal colour, same dotted pattern, same dot spacing, same dot size. The viewer must NOT be able to tell which curve is the query just by looking at the lines themselves; the only distinguishing feature allowed is the literal text 'Start' label sitting on the query curve. Trace each source curve EXACTLY (do NOT smooth, shorten, reroute, or break any curve). • Each single-letter region label stays at the EXACT same position with the SAME glyph, drawn in a small painted ribbon banner.\n\nNo new regions, no new paths, no new labels. No photorealism, no 3D perspective. Pure flat storybook watercolour look. Do not colour-code or style-code the paths differently from each other.\n\nThe source image contains a literal 'Start' text label inside a small white box with a thin black border, placed at the beginning of the query curve. PRESERVE THIS 'Start' LABEL EXACTLY AS-IS in the output: same wording (\"Start\"), same position, same small white box with thin black border, same legibility. Do NOT replace the label with any icon, do NOT translate the word, do NOT restyle the box. The 'Start' label is the ONLY distinguishing mark between the query curve and the distractor curves; every curve must otherwise look identical (same color, same dash pattern, same weight, same texture).", "Re-render as a flat vintage transit map (think: classic mid-century subway diagram printed on cream paper). The output should look LIKE A PRINTED MAP — flat colour districts, clean borders, decorative typography — not a photorealistic aerial photo. Stylisation makes precise geometry easier to preserve.\n\nGEOMETRY MUST BE PRESERVED EXACTLY:\n • Each coloured region stays at the EXACT same shape, position, and adjacency. Replace each region's flat colour with ONE solid printed-ink district fill (mauve, ochre, mint, rose, navy, sand, slate, terracotta) — choose ONE colour per region; do NOT mix colours inside a region; do NOT bleed colour across a boundary.\n • Inside each district, scatter 3–6 tiny printed landmark icons appropriate to a transit map (small church spire, tree, square dot for park, fountain symbol, monument, library glyph, fountain, plus-sign for hospital, train station icon). Vary icons across districts so each district has a slightly different mix of small landmarks. NEVER let an icon bleed across a boundary.\n • Boundaries between regions are thin matte black printed-ink lines, traced EXACTLY along the source partition curves.\n • Every black curve in the source becomes a SINGLE UNIFORM-STYLE printed metro line in the output. ALL curves — the query curve AND every distractor curve — must be rendered IDENTICALLY: same dark-grey ink colour, same uniform thin solid stroke, same weight, same texture. The viewer must NOT be able to tell which curve is the query just by looking at the lines themselves; the only distinguishing feature allowed is the literal text 'Start' label sitting on the query curve. Trace each source curve EXACTLY (do NOT smooth, shorten, reroute, or break any curve).\n • Each single-letter region label stays at the EXACT same position with the SAME glyph, set in a small printed serif tag.\n\nNo new regions, no new lines, no new labels. No photorealism, no 3D perspective. Pure flat printed transit-map look. Do not colour-code or style-code the lines differently from each other.\n\nThe source image contains a literal 'Start' text label inside a small white box with a thin black border, placed at the beginning of the query curve. PRESERVE THIS 'Start' LABEL EXACTLY AS-IS in the output: same wording (\"Start\"), same position, same small white box with thin black border, same legibility. Do NOT replace the label with any icon, do NOT translate the word, do NOT restyle the box. The 'Start' label is the ONLY distinguishing mark between the query curve and the distractor curves; every curve must otherwise look identical (same color, same dash pattern, same weight, same texture).", "Re-render as a flat architectural floor plan (think: clean blueprint or museum gallery plan, drawn on light-grey paper). The output should look LIKE A DRAFTED PLAN — flat coloured rooms, fine pen borders, technical clarity — not a photorealistic aerial photo. Stylisation makes precise geometry easier to preserve.\n\nGEOMETRY MUST BE PRESERVED EXACTLY:\n • Each coloured region stays at the EXACT same shape, position, and adjacency. Replace each region's flat colour with ONE muted architectural fill (parquet wood, polished concrete tone, carpet hatch, terrazzo speckle, marble, wood-deck) — choose ONE fill per region; do NOT mix fills inside a region; do NOT bleed a fill across a boundary.\n • Inside each room, draft 4–8 small simplified-symbol furniture pieces appropriate to a floor plan (rectangle = table or sofa, oval = rug, small circle = chair, square = stool, U-shape = sink, narrow rectangle = bed, slim rectangle = bookcase, dashed-rectangle = doorway). Vary the furniture mix across rooms so each room reads as a different room type. NEVER let a piece of furniture bleed across a boundary.\n • Boundaries between regions are thin uniform-weight black ink lines (the \"wall\" symbol of a floor plan), traced EXACTLY along the source partition curves.\n • Every black curve in the source becomes a SINGLE UNIFORM-STYLE walking-path overlay in the output. ALL curves — the query curve AND every distractor curve — must be rendered IDENTICALLY: same dark-grey colour, same uniform dotted pattern, same dot spacing, same dot size. The viewer must NOT be able to tell which curve is the query just by looking at the lines themselves; the only distinguishing feature allowed is the literal text 'Start' label sitting on the query curve. Trace each source curve EXACTLY (do NOT smooth, shorten, reroute, or break any curve).\n • Each single-letter region label stays at the EXACT same position with the SAME glyph, set in a small drafted-style label box.\n\nNo new regions, no new paths, no new labels. No photorealism, no 3D perspective. Pure flat floor-plan look. Do not colour-code or style-code the paths differently from each other.\n\nThe source image contains a literal 'Start' text label inside a small white box with a thin black border, placed at the beginning of the query curve. PRESERVE THIS 'Start' LABEL EXACTLY AS-IS in the output: same wording (\"Start\"), same position, same small white box with thin black border, same legibility. Do NOT replace the label with any icon, do NOT translate the word, do NOT restyle the box. The 'Start' label is the ONLY distinguishing mark between the query curve and the distractor curves; every curve must otherwise look identical (same color, same dash pattern, same weight, same texture)." ] }, "sequential_traversal/line_intersections": { "src": "images/line_intersections_00000.png", "prompts": [ "Re-render as a photorealistic top-down photograph of curving strands on damp wet sand at low tide. ALL strands are SAME uniform colour and material (e.g. dark green-brown braided seaweed, OR pale bleached driftwood twine, OR crimson kelp fronds — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under (one strand clearly on top of the other), never fuse or branch. NOISY surface: shell fragments, scattered pebbles, ribbons of foam, faint footprints, tide-line debris, ripple marks; the strands sit visibly on top and are NEVER hidden by clutter. Soft directional morning light, real shadows, slight specular wet-sand sheen. Each endpoint label (A, B, C, …) stays at the EXACT same position with the SAME letter glyph, fully readable. No text other than these endpoint labels. No overlays.", "Re-render as a photorealistic top-down photograph of curving cables on a worn concrete workshop floor. ALL strands are SAME uniform colour and material (e.g. bright orange paracord, OR matte black rubber cable, OR coiled copper wire — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under, never fuse or branch. NOISY surface: scattered sawdust, metal shavings, rust patches, oil stains, dropped screws, gaffer tape scraps; the cables sit visibly on top, never hidden. Hard directional shop-lamp light, sharp shadows, slight grease sheen on the cables. Each endpoint label (A, B, C, …) stays at the EXACT same position with the SAME letter glyph, fully readable. No text other than these endpoint labels. No overlays.", "Re-render as a photorealistic top-down photograph of curving slender objects on a damp dappled forest floor. ALL strands are SAME uniform colour and material (e.g. weathered grey twine, OR thin twisted vine stalks, OR pale tan jute rope — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under, never fuse or branch. NOISY surface: fallen leaves of mixed colour, pine needles, twigs, moss patches, damp soil with bark fragments; the strands sit visibly on top, never hidden. Dappled sunlight through canopy gives uneven highlights and soft shadows. Each endpoint label (A, B, C, …) stays at the EXACT same position with the SAME letter glyph, fully readable. No text other than these endpoint labels. No overlays.", "Re-render as a photorealistic top-down photograph of curving garden hoses lying on a sunlit patio. ALL strands are SAME uniform colour and material (e.g. dark-green rubber hose, OR yellow-coiled garden hose — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under crossing (one strand clearly on top of the other), never fuse or branch. NOISY surface: scattered leaves, pebbles, watering cans on the edge, soil flecks, water drops, faint footprints; the hoses sit visibly on top, never hidden. Soft afternoon sunlight, real shadows, slight rubber sheen. Each endpoint label (A, B, C, …) stays at the EXACT same position with the SAME letter glyph, fully readable. No text other than these endpoint labels. No overlays.", "Re-render as a photorealistic top-down photograph of curving ivy vines lying on a damp moss-covered stone path. ALL strands are SAME uniform colour and material (e.g. fresh dark-green ivy vines, OR weathered grey-brown bramble shoots — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under (one vine clearly resting on top of the other), never fuse or branch. NOISY surface: damp stone slabs, scattered fallen leaves, moss patches, twigs, small puddles, lichen flecks; the vines sit visibly on top, never hidden. Soft directional daylight, real shadows, slight wet sheen on the stone. Each endpoint label (A, B, C, …) stays at the EXACT same position with the SAME letter glyph, fully readable. No text other than these endpoint labels. No overlays." ] }, "sequential_traversal/maze": { "src": "images/maze_00000.png", "prompts": [ "Re-render as a photorealistic top-down aerial photograph of a real green hedge maze in a country garden. Every wall in the source becomes a thick well-trimmed hedge tracing the EXACT same path; every walkable corridor becomes mown grass.\n\nCRITICAL — preserve every border opening:\n • The maze has EXACTLY {n} labeled openings on its border, with letters: {labels}. Each opening is a clear gap in the perimeter hedge that allows entry/exit at the EXACT same border position as in the source. The output MUST contain ALL {n} openings in their EXACT source positions, each clearly marked by its corresponding letter ({labels}). Do NOT miss, merge, relocate, add, or rename any opening.\n • Each letter is rendered next to its opening as a small wooden signpost so the letter is unambiguously readable.\n\nWalls (hedges) follow source paths exactly — never add, remove, thicken beyond uniform width, or break a wall. No new openings anywhere except those {n} listed. No people, no animals, no fountains, no decorative plantings inside the corridors. No text other than the {n} letters {labels}.", "Re-render as a photorealistic top-down aerial photograph of a stone-walled medieval castle dungeon, lit by torches mounted along the walls. Every wall in the source becomes a thick weathered grey-stone block wall tracing the EXACT same path; every walkable corridor becomes a flagstone floor with subtle wear.\n\nCRITICAL — preserve every border opening:\n • The dungeon has EXACTLY {n} labeled archway openings on its outer wall, with letters: {labels}. Each archway is a clear breach in the perimeter stone wall at the EXACT same border position as in the source. The output MUST contain ALL {n} archways in their EXACT source positions, each clearly marked by its corresponding letter ({labels}). Do NOT miss, merge, relocate, add, or rename any archway.\n • Each letter is rendered above its archway as a small carved-stone keystone plaque so the letter is unambiguously readable.\n\nWalls (stone blocks) follow source paths exactly — never add, remove, thicken, or break a wall. No new archways anywhere except those {n} listed. No furniture, treasure chests, or props inside the corridors. No text other than the {n} letters {labels}.\n\n⚠️ WALL-PRESENCE PRECISION (HIGHEST PRIORITY): EVERY single wall segment in the source MUST be present in the output — including short stubs, dead-ends, T-junctions, and isolated single-cell walls. Do NOT skip, omit, shorten, merge, fuse, or thicken-away any wall. Every wall pixel-path in the source must have a corresponding wall in the rendered scene. A wall that exists in the source but is missing or smoothed-over in the output is a critical failure. Equally: do NOT add any new wall not present in the source. The maze WALL TOPOLOGY must be identical to the source within ~3 px on every wall path.", "Re-render as a photorealistic top-down aerial photograph of a real autumn cornfield maze cut into a working farm field. Every wall in the source becomes a tall standing block of corn stalks tracing the EXACT same path; every walkable corridor becomes a flat cleared dirt path through the field.\n\nCRITICAL — preserve every border opening:\n • The cornfield maze has EXACTLY {n} labeled border openings, with letters: {labels}. Each opening is a clear gap in the perimeter corn block that allows entry/exit at the EXACT same border position as in the source. The output MUST contain ALL {n} openings in their EXACT source positions, each clearly marked by its corresponding letter ({labels}). Do NOT miss, merge, relocate, add, or rename any opening.\n • Each letter is rendered next to its opening as a painted wooden farm sign so the letter is unambiguously readable.\n\nWalls (corn stalks) follow source paths exactly — never add, remove, thicken beyond uniform width, or break a wall. No new openings anywhere except those {n} listed. No farm equipment, scarecrows, or visitors inside the corridors. No text other than the {n} letters {labels}.\n\n⚠️ WALL-PRESENCE PRECISION (HIGHEST PRIORITY): EVERY single wall segment in the source MUST be present in the output — including short stubs, dead-ends, T-junctions, and isolated single-cell walls. Do NOT skip, omit, shorten, merge, fuse, or thicken-away any wall. Every wall pixel-path in the source must have a corresponding wall in the rendered scene. A wall that exists in the source but is missing or smoothed-over in the output is a critical failure. Equally: do NOT add any new wall not present in the source. The maze WALL TOPOLOGY must be identical to the source within ~3 px on every wall path.", "Re-render as a photorealistic top-down aerial photograph of a vast library seen from directly above. Every wall in the source becomes a tall double-sided wooden bookshelf packed with books, tracing the EXACT same path; every walkable corridor becomes a polished hardwood floor.\n\nCRITICAL — preserve every border opening:\n • The library has EXACTLY {n} labeled border openings, with letters: {labels}. Each opening is a clear gap in the perimeter shelving at the EXACT same border position as in the source. The output MUST contain ALL {n} openings in their EXACT source positions, each clearly marked by its corresponding letter ({labels}). Do NOT miss, merge, relocate, add, or rename any opening.\n • Each letter is rendered above its opening as a small brass directory plate so the letter is unambiguously readable.\n\nWalls (bookshelves) follow source paths exactly — never add, remove, thicken beyond uniform width, or break a wall. No new openings anywhere except those {n} listed. No people, no chairs, no reading lamps, no carts inside the corridors. No text other than the {n} letters {labels}.\n\n⚠️ WALL-PRESENCE PRECISION (HIGHEST PRIORITY): EVERY single wall segment in the source MUST be present in the output — including short stubs, dead-ends, T-junctions, and isolated single-cell walls. Do NOT skip, omit, shorten, merge, fuse, or thicken-away any wall. Every wall pixel-path in the source must have a corresponding wall in the rendered scene. A wall that exists in the source but is missing or smoothed-over in the output is a critical failure. Equally: do NOT add any new wall not present in the source. The maze WALL TOPOLOGY must be identical to the source within ~3 px on every wall path.", "Re-render as a photorealistic top-down aerial photograph of a snow-path maze cut into deep arctic snowfields. Every wall in the source becomes a tall packed snowbank tracing the EXACT same path; every walkable corridor becomes pressed flat snow with subtle shoe scuffs.\n\nCRITICAL — preserve every border opening:\n • The maze has EXACTLY {n} labeled border openings, with letters: {labels}. Each opening is a clear gap in the perimeter snowbank at the EXACT same border position as in the source. The output MUST contain ALL {n} openings in their EXACT source positions, each clearly marked by its corresponding letter ({labels}). Do NOT miss, merge, relocate, add, or rename any opening.\n • Each letter is rendered next to its opening as a small wooden trail-marker post so the letter is unambiguously readable.\n\nWalls (snowbanks) follow source paths exactly — never add, remove, thicken beyond uniform width, or break a wall. No new openings anywhere except those {n} listed. No people, no sleds, no animals, no equipment inside the corridors. No text other than the {n} letters {labels}." ] }, "sequential_traversal/traverse_ordering": { "src": "images/traverse_ordering_00000.png", "prompts": [ "Re-render as a photorealistic top-down photograph of curving strands on damp wet sand at low tide. ALL strands are SAME uniform colour and material (e.g. dark green-brown braided seaweed, OR pale bleached driftwood twine, OR crimson kelp fronds — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under (one strand clearly on top of the other), never fuse or branch. NOISY surface: shell fragments, scattered pebbles, ribbons of foam, faint footprints, tide-line debris, ripple marks; the strands sit visibly on top and are NEVER hidden by clutter. Soft directional morning light, real shadows, slight specular wet-sand sheen. Every labelled crossing point and the green 'S' start marker stay at the EXACT same position with the SAME letter glyph and the SAME green colour for 'S'. No text other than these existing labels and the 'S'. No overlays.", "Re-render as a photorealistic top-down photograph of curving cables on a worn concrete workshop floor. ALL strands are SAME uniform colour and material (e.g. bright orange paracord, OR matte black rubber cable, OR coiled copper wire — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under, never fuse or branch. NOISY surface: scattered sawdust, metal shavings, rust patches, oil stains, dropped screws, gaffer tape scraps; the cables sit visibly on top, never hidden. Hard directional shop-lamp light, sharp shadows, slight grease sheen on the cables. Every labelled crossing point and the green 'S' start marker stay at the EXACT same position with the SAME letter glyph and the SAME green colour for 'S'. No text other than these existing labels and the 'S'. No overlays.", "Re-render as a photorealistic top-down photograph of curving slender objects on a damp dappled forest floor. ALL strands are SAME uniform colour and material (e.g. weathered grey twine, OR thin twisted vine stalks, OR pale tan jute rope — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under, never fuse or branch. NOISY surface: fallen leaves of mixed colour, pine needles, twigs, moss patches, damp soil with bark fragments; the strands sit visibly on top, never hidden. Dappled sunlight through canopy gives uneven highlights and soft shadows. Every labelled crossing point and the green 'S' start marker stay at the EXACT same position with the SAME letter glyph and the SAME green colour for 'S'. No text other than these existing labels and the 'S'. No overlays.", "Re-render as a photorealistic top-down photograph of curving garden hoses lying on a sunlit patio. ALL strands are SAME uniform colour and material (e.g. dark-green rubber hose, OR yellow-coiled garden hose — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under, never fuse or branch. NOISY surface: scattered leaves, pebbles, watering cans on the edge, soil flecks, water drops, faint footprints; the hoses sit visibly on top, never hidden. Soft afternoon sunlight, real shadows, slight rubber sheen. Every labelled crossing point and the green 'S' start marker stay at the EXACT same position with the SAME letter glyph and the SAME green colour for 'S'. No text other than these existing labels and the 'S'. No overlays.", "Re-render as a photorealistic top-down photograph of curving ivy vines lying on a damp moss-covered stone path. ALL strands are SAME uniform colour and material (e.g. fresh dark-green ivy vines, OR weathered grey-brown bramble shoots — pick ONE per image, same across every strand). Each strand follows the source path EXACTLY; at every crossing, render an unambiguous over/under (one vine clearly resting on top of the other), never fuse or branch. NOISY surface: damp stone slabs, scattered fallen leaves, moss patches, twigs, small puddles, lichen flecks; the vines sit visibly on top, never hidden. Soft directional daylight, real shadows, slight wet sheen on the stone. Every labelled crossing point and the green 'S' start marker stay at the EXACT same position with the SAME letter glyph and the SAME green colour for 'S'. No text other than these existing labels and the 'S'. No overlays." ] }, "visual_attribute_transfer/constellation_match_count": { "src": "images/constellation_match_count_00000.png", "prompts": [ "Re-render as a deep-space astrophotograph. Preserve EXACTLY: the LEFT 'TEMPLATE' bordered box (with its label and frame), the vertical divider, and the position of every white dot on both sides (within ~2 px).\n\nLEFT (template): each dot → one identical small white star, all matching, on a uniform dark sky.\n\nRIGHT (field): each dot → ONE star, but EVERY star is rendered DIFFERENTLY — vary colour (white, blue, yellow, orange, red), brightness, diffraction-spike pattern, halo, hue, faint nebula tint. Sizes vary slightly but stay small (similar order to template stars). Positions must be pixel-exact — no shifting, no adding, no removing.\n\nNo text overlays beyond the existing TEMPLATE label.", "Re-render as a top-down photograph on black velvet. Preserve EXACTLY: the LEFT 'TEMPLATE' bordered box (with its label and frame), the vertical divider, and the position of every dot on both sides (within ~2 px).\n\nLEFT (template): each dot → one identical small white pearl, all matching.\n\nRIGHT (field): each dot → ONE small gemstone, but EVERY stone is rendered DIFFERENTLY — vary colour (ruby, sapphire, emerald, citrine, amethyst, opal, pearl, garnet), cut (round, oval, faceted, cabochon), material (gem, polished bead, glass), and surface sheen. Keep all stones small (no larger than ~1.5× a template pearl). Positions must be pixel-exact — no shifting, no adding, no removing.\n\nNo text overlays beyond the existing TEMPLATE label.", "Re-render as a top-down photograph at night. Preserve EXACTLY: the LEFT 'TEMPLATE' bordered box (with its label and frame), the vertical divider, and the position of every dot on both sides (within ~2 px).\n\nLEFT (template): each dot → one identical small warm tealight flame, all matching, on dark wood.\n\nRIGHT (field): each dot → ONE point of light, but EVERY light is rendered DIFFERENTLY — vary type (candle flame, lantern bulb, fairy LED, bokeh glint, fibre-optic tip, glowstick dot), colour (warm white, gold, blue, green, magenta, red), bloom shape, and halo size. Keep each light small (no larger than ~1.5× a template flame). Positions must be pixel-exact — no shifting, no adding, no removing.\n\nNo text overlays beyond the existing TEMPLATE label.", "Re-render as a top-down photograph on a dark slate plate. Preserve EXACTLY: the LEFT 'TEMPLATE' bordered box (with its label and frame), the vertical divider, and the position of every dot on both sides (within ~2 px).\n\nLEFT (template): each dot → one identical small white peppermint candy, all matching.\n\nRIGHT (field): each dot → ONE small confection, but EVERY one is rendered DIFFERENTLY — vary type (gumdrop, jelly bean, hard candy, chocolate ball, marshmallow, pill, dragee, m&m), colour, gloss, texture (smooth, sugar-dusted, crystallised, matte). Keep each piece small (no larger than ~1.5× a template mint). Positions must be pixel-exact — no shifting, no adding, no removing.\n\nNo text overlays beyond the existing TEMPLATE label.", "Re-render as a top-down photograph on wet dark sand. Preserve EXACTLY: the LEFT 'TEMPLATE' bordered box (with its label and frame), the vertical divider, and the position of every dot on both sides (within ~2 px).\n\nLEFT (template): each dot → one identical small white shell, all matching.\n\nRIGHT (field): each dot → ONE small sea object, but EVERY one is rendered DIFFERENTLY — vary type (cowrie, scallop, cone shell, sea-glass chip, polished pebble, coral fragment, starfish point, mother-of-pearl chip), colour, texture (glossy, ridged, matte, translucent), and material. Keep each object small (no larger than ~1.5× a template shell). Positions must be pixel-exact — no shifting, no adding, no removing.\n\nNo text overlays beyond the existing TEMPLATE label." ] }, "visual_attribute_transfer/contour_silhouette_count": { "src": "images/contour_silhouette_count_00000.png", "prompts": [ "GEOMETRY IS A PRECISION TEMPLATE — DO NOT CHANGE IT. Re-render as a photorealistic top-down photograph. The source defines exact pixel positions and sizes; copy them.\n\n⚠️ PIXEL-EXACT (within ~2 px, NON-NEGOTIABLE):\n • LEFT TEMPLATE SIZE & POSITION: the single contour keeps its EXACT bounding box (width × height) and EXACT (x, y) from the source. If the source template is small, render it small. Do NOT enlarge, do NOT recentre, do NOT zoom in, do NOT make it fill the LEFT panel.\n • RIGHT FIELD SIZE & POSITION: every source contour → ONE object at the same bounding box and same (x, y). Same count as source — never add, skip, merge, or split.\n • RELATIVE SIZES: template-vs-field and field-vs-field ratios match the source pixel-for-pixel.\n • DIVIDER: vertical separator at the source x-coordinate; LEFT/RIGHT panel widths match the source.\n\nSTRUCTURE: LEFT → ONE class-A object on LEFT surface. RIGHT → many class-B objects on RIGHT surface.\n\nPick TWO different items — one for LEFT (class A), a different one for RIGHT (class B):\n pebbles, river stones, autumn leaves, pottery shards, mushrooms, bread rolls, cookies, pancakes, pastries, dumplings, silk pouches, beach shells, driftwood pieces, clay tiles, wax seals, sponge pieces.\n\nLEFT surface: worn wooden cutting board. RIGHT surface: moss-covered stone slab. Each panel has its own directional lighting. Per-object micro-variation in shade/sheen. Background must not create new object-like blobs.\n\nNo text, no labels, no overlays.", "⚠️ CRITICAL CONSTRAINTS — HIGHEST PRIORITY (do not violate):\n 1. SCALE IS LOCKED. Every rendered object must occupy the EXACT same pixel area as its source contour/stroke. Do NOT enlarge, shrink, or normalise. The single template object on the LEFT must match the source template's pixel size exactly. Each field object on the RIGHT must match the pixel size of its corresponding source contour/stroke exactly. Relative sizes between all objects (left vs right and right-vs-right) must be preserved within ~5 px. The model must NOT make the LEFT template object larger or smaller than its source, and must NOT rescale the RIGHT field objects to fill space.\n 2. DIVIDER POSITION IS LOCKED. The vertical separator line between LEFT and RIGHT panels must be at the EXACT same x-coordinate as in the source image. Do NOT shift, widen, narrow, or recentre the divider. The LEFT panel pixel width and RIGHT panel pixel width must each match the source exactly. The single template object on the LEFT must remain at the SAME (x, y) position as the source template stroke/contour — do NOT recentre it within the LEFT panel.\n\nRe-render the two panels as a photorealistic top-down photograph. The two panels are separated by the SAME vertical divider, share the SAME overall lighting bias, but use TWO DIFFERENT object types and TWO DIFFERENT surfaces.\n\nPANEL STRUCTURE — read carefully:\n • The LEFT panel is the TEMPLATE: it contains EXACTLY ONE closed contour in the source. Render this single template contour as ONE object of class A on the LEFT surface, preserving the EXACT silhouette outline, size, and position of that single contour.\n • The RIGHT panel is the FIELD: it contains MANY closed contours in the source. Render EVERY contour in the right-panel source as ONE object of class B on the RIGHT surface, preserving the EXACT silhouette outline, size, and position of each contour. The number of objects in the RIGHT panel must equal the source's count exactly.\n\nPick TWO different items at random from this pool — one for the LEFT (class A, the single template) and a different one for the RIGHT (class B, the many field objects):\n vintage coins, brass buttons, sea glass, polished gemstones, wax seals, glass beads, metal tokens, bottle caps, mosaic chips, glass marbles, mother-of-pearl discs, enamel buttons.\n\nThe LEFT surface is a velvet jeweller's tray; the RIGHT surface is a slate counter top. Each panel has its own directional lighting (slightly different angle / colour temperature).\n\nACCURACY IS CRITICAL:\n • LEFT panel: exactly ONE class-A object whose silhouette matches the single template contour outline within ~5 px. No drift, no rescale, no rotation.\n • RIGHT panel: EVERY source contour becomes exactly ONE class-B object whose silhouette matches its source contour within ~5 px. Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each rendered object's silhouette is identical to its source contour. A silhouette overlay should align almost perfectly within each panel.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "⚠️ CRITICAL CONSTRAINTS — HIGHEST PRIORITY (do not violate):\n 1. SCALE IS LOCKED. Every rendered object must occupy the EXACT same pixel area as its source contour/stroke. Do NOT enlarge, shrink, or normalise. The single template object on the LEFT must match the source template's pixel size exactly. Each field object on the RIGHT must match the pixel size of its corresponding source contour/stroke exactly. Relative sizes between all objects (left vs right and right-vs-right) must be preserved within ~5 px. The model must NOT make the LEFT template object larger or smaller than its source, and must NOT rescale the RIGHT field objects to fill space.\n 2. DIVIDER POSITION IS LOCKED. The vertical separator line between LEFT and RIGHT panels must be at the EXACT same x-coordinate as in the source image. Do NOT shift, widen, narrow, or recentre the divider. The LEFT panel pixel width and RIGHT panel pixel width must each match the source exactly. The single template object on the LEFT must remain at the SAME (x, y) position as the source template stroke/contour — do NOT recentre it within the LEFT panel.\n\nRe-render the two panels as a photorealistic top-down photograph. The two panels are separated by the SAME vertical divider, share the SAME overall lighting bias, but use TWO DIFFERENT object types and TWO DIFFERENT surfaces.\n\nPANEL STRUCTURE — read carefully:\n • The LEFT panel is the TEMPLATE: it contains EXACTLY ONE closed contour in the source. Render this single template contour as ONE object of class A on the LEFT surface, preserving the EXACT silhouette outline, size, and position of that single contour.\n • The RIGHT panel is the FIELD: it contains MANY closed contours in the source. Render EVERY contour in the right-panel source as ONE object of class B on the RIGHT surface, preserving the EXACT silhouette outline, size, and position of each contour. The number of objects in the RIGHT panel must equal the source's count exactly.\n\nPick TWO different items at random from this pool — one for the LEFT (class A, the single template) and a different one for the RIGHT (class B, the many field objects):\n grapes, cherries, olives, dried apricots, macarons, chocolate truffles, hard candies, dried figs, kumquats, lychees, almonds, blueberries, dried cranberries, raisins, gummies.\n\nThe LEFT surface is a marble pastry slab; the RIGHT surface is a raw linen tablecloth. Each panel has its own directional lighting (slightly different angle / colour temperature).\n\nACCURACY IS CRITICAL:\n • LEFT panel: exactly ONE class-A object whose silhouette matches the single template contour outline within ~5 px. No drift, no rescale, no rotation.\n • RIGHT panel: EVERY source contour becomes exactly ONE class-B object whose silhouette matches its source contour within ~5 px. Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each rendered object's silhouette is identical to its source contour. A silhouette overlay should align almost perfectly within each panel.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "⚠️ CRITICAL CONSTRAINTS — HIGHEST PRIORITY (do not violate):\n 1. SCALE IS LOCKED. Every rendered object must occupy the EXACT same pixel area as its source contour/stroke. Do NOT enlarge, shrink, or normalise. The single template object on the LEFT must match the source template's pixel size exactly. Each field object on the RIGHT must match the pixel size of its corresponding source contour/stroke exactly. Relative sizes between all objects (left vs right and right-vs-right) must be preserved within ~5 px. The model must NOT make the LEFT template object larger or smaller than its source, and must NOT rescale the RIGHT field objects to fill space.\n 2. DIVIDER POSITION IS LOCKED. The vertical separator line between LEFT and RIGHT panels must be at the EXACT same x-coordinate as in the source image. Do NOT shift, widen, narrow, or recentre the divider. The LEFT panel pixel width and RIGHT panel pixel width must each match the source exactly. The single template object on the LEFT must remain at the SAME (x, y) position as the source template stroke/contour — do NOT recentre it within the LEFT panel.\n\nRe-render the two panels as a photorealistic top-down photograph. The two panels are separated by the SAME vertical divider, share the SAME overall lighting bias, but use TWO DIFFERENT object types and TWO DIFFERENT surfaces.\n\nPANEL STRUCTURE — read carefully:\n • The LEFT panel is the TEMPLATE: it contains EXACTLY ONE closed contour in the source. Render this single template contour as ONE object of class A on the LEFT surface, preserving the EXACT silhouette outline, size, and position of that single contour.\n • The RIGHT panel is the FIELD: it contains MANY closed contours in the source. Render EVERY contour in the right-panel source as ONE object of class B on the RIGHT surface, preserving the EXACT silhouette outline, size, and position of each contour. The number of objects in the RIGHT panel must equal the source's count exactly.\n\nPick TWO different items at random from this pool — one for the LEFT (class A, the single template) and a different one for the RIGHT (class B, the many field objects):\n rubber washers, metal nuts, plastic discs, silicone gaskets, marble cabochons, wooden game pieces, cardboard punch-outs, ceramic tiles, linoleum scraps, cork rounds, leather punches, felted-wool dots.\n\nThe LEFT surface is a galvanised steel workbench; the RIGHT surface is a weathered concrete floor. Each panel has its own directional lighting (slightly different angle / colour temperature).\n\nACCURACY IS CRITICAL:\n • LEFT panel: exactly ONE class-A object whose silhouette matches the single template contour outline within ~5 px. No drift, no rescale, no rotation.\n • RIGHT panel: EVERY source contour becomes exactly ONE class-B object whose silhouette matches its source contour within ~5 px. Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each rendered object's silhouette is identical to its source contour. A silhouette overlay should align almost perfectly within each panel.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "⚠️ CRITICAL CONSTRAINTS — HIGHEST PRIORITY (do not violate):\n 1. SCALE IS LOCKED. Every rendered object must occupy the EXACT same pixel area as its source contour/stroke. Do NOT enlarge, shrink, or normalise. The single template object on the LEFT must match the source template's pixel size exactly. Each field object on the RIGHT must match the pixel size of its corresponding source contour/stroke exactly. Relative sizes between all objects (left vs right and right-vs-right) must be preserved within ~5 px. The model must NOT make the LEFT template object larger or smaller than its source, and must NOT rescale the RIGHT field objects to fill space.\n 2. DIVIDER POSITION IS LOCKED. The vertical separator line between LEFT and RIGHT panels must be at the EXACT same x-coordinate as in the source image. Do NOT shift, widen, narrow, or recentre the divider. The LEFT panel pixel width and RIGHT panel pixel width must each match the source exactly. The single template object on the LEFT must remain at the SAME (x, y) position as the source template stroke/contour — do NOT recentre it within the LEFT panel.\n\nRe-render the two panels as a photorealistic top-down photograph. The two panels are separated by the SAME vertical divider, share the SAME overall lighting bias, but use TWO DIFFERENT object types and TWO DIFFERENT surfaces.\n\nPANEL STRUCTURE — read carefully:\n • The LEFT panel is the TEMPLATE: it contains EXACTLY ONE closed contour in the source. Render this single template contour as ONE object of class A on the LEFT surface, preserving the EXACT silhouette outline, size, and position of that single contour.\n • The RIGHT panel is the FIELD: it contains MANY closed contours in the source. Render EVERY contour in the right-panel source as ONE object of class B on the RIGHT surface, preserving the EXACT silhouette outline, size, and position of each contour. The number of objects in the RIGHT panel must equal the source's count exactly.\n\nPick TWO different items at random from this pool — one for the LEFT (class A, the single template) and a different one for the RIGHT (class B, the many field objects):\n silk pouches, fabric patches, felt cutouts, paper origami pieces, leather scraps, lace pieces, suede swatches, linen squares, vellum cutouts, cork punches, foam-board chips.\n\nThe LEFT surface is a craft-paper sheet with creases; the RIGHT surface is a denim-blue cotton sheet. Each panel has its own directional lighting (slightly different angle / colour temperature).\n\nACCURACY IS CRITICAL:\n • LEFT panel: exactly ONE class-A object whose silhouette matches the single template contour outline within ~5 px. No drift, no rescale, no rotation.\n • RIGHT panel: EVERY source contour becomes exactly ONE class-B object whose silhouette matches its source contour within ~5 px. Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each rendered object's silhouette is identical to its source contour. A silhouette overlay should align almost perfectly within each panel.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays." ] }, "visual_attribute_transfer/spot_the_contour_diff": { "src": "images/contour_diff_00000.png", "prompts": [ "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool, one for the LEFT panel and a different one for the RIGHT panel:\n pebbles, river stones, autumn leaves, pottery shards, mushrooms, bread rolls, cookies, pancakes, pastries, dumplings, silk pouches, leather scraps, fabric patches, beach shells, driftwood pieces, clay tiles, wax seals, sponge pieces.\n\nThe LEFT panel sits on a worn wooden cutting board; the RIGHT panel sits on a moss-covered stone slab. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY closed contour in the source becomes EXACTLY ONE object on the matching side, at the SAME silhouette outline, SAME size, SAME (x, y) — within ~5 px. No drift, no rescale, no rotation away from the source.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each silhouette is identical to the source contour on its side. A silhouette overlay should align almost perfectly.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same silhouette renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool, one for the LEFT panel and a different one for the RIGHT panel:\n vintage coins, brass buttons, sea glass, polished gemstones, wax seals, glass beads, metal tokens, bottle caps, mosaic chips, glass marbles, mother-of-pearl discs, enamel buttons.\n\nThe LEFT panel sits on a velvet jeweller's tray; the RIGHT panel sits on a slate counter top. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY closed contour in the source becomes EXACTLY ONE object on the matching side, at the SAME silhouette outline, SAME size, SAME (x, y) — within ~5 px. No drift, no rescale, no rotation away from the source.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each silhouette is identical to the source contour on its side. A silhouette overlay should align almost perfectly.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same silhouette renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool, one for the LEFT panel and a different one for the RIGHT panel:\n grapes, cherries, olives, dried apricots, macarons, chocolate truffles, hard candies, dried figs, kumquats, lychees, almonds, blueberries, dried cranberries, raisins, gummies.\n\nThe LEFT panel sits on a marble pastry slab; the RIGHT panel sits on a raw linen tablecloth. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY closed contour in the source becomes EXACTLY ONE object on the matching side, at the SAME silhouette outline, SAME size, SAME (x, y) — within ~5 px. No drift, no rescale, no rotation away from the source.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each silhouette is identical to the source contour on its side. A silhouette overlay should align almost perfectly.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same silhouette renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool, one for the LEFT panel and a different one for the RIGHT panel:\n rubber washers, metal nuts, plastic discs, silicone gaskets, marble cabochons, wooden game pieces, cardboard punch-outs, ceramic tiles, linoleum scraps, cork rounds, leather punches, felted-wool dots.\n\nThe LEFT panel sits on a galvanised steel workbench; the RIGHT panel sits on a weathered concrete floor. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY closed contour in the source becomes EXACTLY ONE object on the matching side, at the SAME silhouette outline, SAME size, SAME (x, y) — within ~5 px. No drift, no rescale, no rotation away from the source.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each silhouette is identical to the source contour on its side. A silhouette overlay should align almost perfectly.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same silhouette renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool, one for the LEFT panel and a different one for the RIGHT panel:\n silk pouches, fabric patches, felt cutouts, paper origami pieces, leather scraps, lace pieces, suede swatches, linen squares, vellum cutouts, cork punches, foam-board chips.\n\nThe LEFT panel sits on a craft-paper sheet with creases; the RIGHT panel sits on a denim-blue cotton sheet. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY closed contour in the source becomes EXACTLY ONE object on the matching side, at the SAME silhouette outline, SAME size, SAME (x, y) — within ~5 px. No drift, no rescale, no rotation away from the source.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split contours.\n • Object TYPE differs left vs right, but each silhouette is identical to the source contour on its side. A silhouette overlay should align almost perfectly.\n • Background fills empty space only — must not create new object-like blobs.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same silhouette renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding objects across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays." ] }, "visual_attribute_transfer/spot_the_field_diff": { "src": "images/field_diff_00000.png", "prompts": [ "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every cell value, every cell colour, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-panel sheet pinned to a corkboard. Add visible cork texture surrounding the sheet, two coloured push-pins at the top corners, and gentle warm lighting from a desk lamp at upper left. The sheet sits flat against the cork; do NOT crop or rotate the panels. Use a 2:1 wide framing.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every cell value, every cell colour, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-panel sheet lying on a worn wooden desk. Surround the sheet with a few real-world objects positioned OUTSIDE the panels: a closed pen, a coffee-ring stain on the wood, scattered paper-clips, and a corner of an open notebook. Lighting is even daylight. Use a 2:1 wide framing.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every cell value, every cell colour, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-panel sheet held against a clipboard, with the metal clip visible at the top edge OUTSIDE the panels. Show the clipboard's plywood grain and a darker shadow under the right edge. Lighting is fluorescent overhead, mildly cool. Use a 2:1 wide framing.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every cell value, every cell colour, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-panel sheet sitting on a darker textile background — a folded grey wool blanket or felt mat. Add a faint vignette and a small folded paper triangle at the top-right corner OUTSIDE the panels. Lighting is overcast daylight from above. Use a 2:1 wide framing.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every cell value, every cell colour, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-panel sheet in a manila folder, with the folder's tan edge visible at the bottom and a small yellow sticky note in the top-left margin OUTSIDE the panels. The sheet sits inside the folder, slightly raised. Lighting is warm desk lamp from upper right. Use a 2:1 wide framing." ] }, "visual_attribute_transfer/spot_the_signal_diff": { "src": "images/signal_diff_00000.png", "prompts": [ "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every waveform sample, every gridline, every axis tick, every label, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-signal-trace panel as if it were a printed lab strip lying on a metal lab bench. Around the strip add: a partial view of an oscilloscope chassis dial at top-right, a coiled BNC cable at bottom-left, and a printed-circuit board corner at lower-right — ALL outside the panel area. Lighting is bright fluorescent. Use the wide 2:1 ratio of the original strip.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every waveform sample, every gridline, every axis tick, every label, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-signal-trace panel pinned to a lab whiteboard. Show whiteboard texture surrounding the strip, two magnetic clips at top corners, and faint marker scribbles in the visible whiteboard area OUTSIDE the panel. Lighting is even cool office daylight. Use a 2:1 ratio.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every waveform sample, every gridline, every axis tick, every label, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-signal-trace panel as a printout sitting on an electronics workbench. Around the printout (NOT on it) add scattered components: a few resistors, a multi-meter probe tip, and a small breadboard at bottom-right. The printout itself is flat and uncropped. Lighting is warm bench-lamp from above. Use a 2:1 ratio.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every waveform sample, every gridline, every axis tick, every label, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-signal-trace panel as if displayed on a flat-panel monitor seen at slight glancing angle (so the bezel and minor screen reflections frame the panel). Critically, the panel content is rendered crisp and unmodified — only the surrounding monitor bezel + faint room reflection is added. Use a 2:1 ratio.", "⚠️ PIXEL-EXACT (NON-NEGOTIABLE):\n • The TWO-PANEL CONTENT is sacred. Every cell, every grid line, every waveform sample, every gridline, every axis tick, every label, every divider, every panel border MUST be reproduced EXACTLY — do NOT alter, recolour, redraw, blur, smooth, re-rasterise, or shift any cell within the panels.\n • The vertical divider between LEFT and RIGHT panels stays at the EXACT same x-coordinate.\n • Object COUNT inside the panels equals the source EXACTLY — never add, skip, merge, or split anything inside a panel.\n • No signal-level modifications inside the panels: no extra noise, paper grain, blur, JPEG artefacts, or colour shifts on the panel pixels (these have already been baked in upstream). The panels are passed through UNCHANGED.\n • You may ONLY add real-world context AROUND the panels (outside their bounding boxes). Inside the panels: ZERO change.\n\nPlace the unchanged two-signal-trace panel as a strip lying on a darker fabric background (denim or felt), with a small ruler placed alongside (parallel, OUTSIDE the panel) and a metal washer near the corner. Lighting is even daylight. Use a 2:1 ratio." ] }, "visual_attribute_transfer/spot_the_stroke_diff": { "src": "images/stroke_diff_00000.png", "prompts": [ "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool of curve-shaped objects, one for the LEFT panel and a different one for the RIGHT panel:\n thin twigs, dry pine needles, blades of grass, fallen leaf-stems, pieces of twine, paracord segments, ribbons, leather thongs, cooked spaghetti, ramen noodles, copper wire snippets, hairpins, pencil shavings, dried herb stalks, fish bones, eel grass.\n\nThe LEFT panel sits on damp moss with twigs; the RIGHT panel sits on a polished slate plate. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY brush stroke in the source becomes EXACTLY ONE elongated object on the matching side, following the SAME curve path, SAME length, SAME (x, y) — within ~5 px. No drift, no rescale, no flip.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split strokes.\n • Object TYPE differs left vs right, but each curve is identical to the source stroke on its side. A curve overlay should align almost perfectly.\n • Background fills empty space only — must not create new stroke-like shapes.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same curve renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding curves across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool of curve-shaped objects, one for the LEFT panel and a different one for the RIGHT panel:\n yarn strands, embroidery thread, wool roving, cotton thread, silk thread, sinew, horsehair, fishing line, dental floss, jute twine, hemp string, linen thread.\n\nThe LEFT panel sits on a tailor's wooden cutting mat; the RIGHT panel sits on a denim cotton sheet. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY brush stroke in the source becomes EXACTLY ONE elongated object on the matching side, following the SAME curve path, SAME length, SAME (x, y) — within ~5 px. No drift, no rescale, no flip.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split strokes.\n • Object TYPE differs left vs right, but each curve is identical to the source stroke on its side. A curve overlay should align almost perfectly.\n • Background fills empty space only — must not create new stroke-like shapes.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same curve renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding curves across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool of curve-shaped objects, one for the LEFT panel and a different one for the RIGHT panel:\n spaghetti, udon, soba, rice vermicelli, lasagna strips, churros, breadsticks, pretzel sticks, liquorice sticks, chow-mein noodles, fettuccine, linguine.\n\nThe LEFT panel sits on a marble pastry slab; the RIGHT panel sits on a dark slate plate. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY brush stroke in the source becomes EXACTLY ONE elongated object on the matching side, following the SAME curve path, SAME length, SAME (x, y) — within ~5 px. No drift, no rescale, no flip.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split strokes.\n • Object TYPE differs left vs right, but each curve is identical to the source stroke on its side. A curve overlay should align almost perfectly.\n • Background fills empty space only — must not create new stroke-like shapes.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same curve renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding curves across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool of curve-shaped objects, one for the LEFT panel and a different one for the RIGHT panel:\n copper wire, aluminium wire, solder strips, steel cable, nylon zip-ties, fibre-optic strands, jumper wires, leather laces, electric ribbon cable, brass rod offcuts.\n\nThe LEFT panel sits on a galvanised steel workbench; the RIGHT panel sits on a weathered concrete floor. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY brush stroke in the source becomes EXACTLY ONE elongated object on the matching side, following the SAME curve path, SAME length, SAME (x, y) — within ~5 px. No drift, no rescale, no flip.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split strokes.\n • Object TYPE differs left vs right, but each curve is identical to the source stroke on its side. A curve overlay should align almost perfectly.\n • Background fills empty space only — must not create new stroke-like shapes.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same curve renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding curves across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays.", "Re-render the two panels as a photorealistic top-down photograph using TWO DIFFERENT object types for left vs right — same vertical divider, same overall lighting bias. Pick TWO different items at random from this pool of curve-shaped objects, one for the LEFT panel and a different one for the RIGHT panel:\n twigs, dried pine needles, fern fronds, vine tendrils, fallen leaf veins, herb stalks, dried wheat stalks, split bamboo strips, dried lavender stems, juniper sprigs.\n\nThe LEFT panel sits on damp dark forest soil; the RIGHT panel sits on a pale weathered birch plank. The two surfaces share the same lighting but their materials and colours are completely different.\n\nACCURACY IS CRITICAL:\n • EVERY brush stroke in the source becomes EXACTLY ONE elongated object on the matching side, following the SAME curve path, SAME length, SAME (x, y) — within ~5 px. No drift, no rescale, no flip.\n • Same OBJECT COUNT per panel as the source. Never add, skip, merge, or split strokes.\n • Object TYPE differs left vs right, but each curve is identical to the source stroke on its side. A curve overlay should align almost perfectly.\n • Background fills empty space only — must not create new stroke-like shapes.\n\nANTI-PROGRAM-COMPARISON HARDENING:\n • Each panel has a NOISY, lived-in surface — micro-scratches, dust, subtle wear, slight colour drift; same genre of noise across both panels but DIFFERENT realisations on each side.\n • Each panel has its own directional lighting (slightly different angle and colour temperature), so the same curve renders with different shadow direction and brightness on the two sides.\n • Per-object micro-variation (texture / shade / sheen) so corresponding curves across panels never share a pixel-level signature.\n\nNo text, no labels, no overlays." ] }, "visual_attribute_transfer/stroke_gesture_count": { "src": "images/stroke_gesture_count_00000.png", "prompts": [ "Re-render as a photorealistic top-down photograph. Same vertical divider as source.\n\n⚠️ LOCKED CONSTRAINTS:\n • Scale: every rendered curve occupies the EXACT same pixel area as its source stroke. No rescaling, no normalising.\n • Divider: vertical separator stays at the source's exact x-coordinate. LEFT and RIGHT panel widths match the source.\n • Position: every curve stays at the same (x, y) and follows the same path/length/curvature within ~5 px.\n\nPANEL STRUCTURE:\n • LEFT (template): the ONE source stroke → ONE elongated object of class A on the LEFT surface.\n • RIGHT (field): every source stroke → ONE elongated object of class B on the RIGHT surface. Object count must equal source exactly — never add, skip, merge, or split.\n\nPick TWO different items from this pool — one for LEFT (class A), a different one for RIGHT (class B):\n thin twigs, dry pine needles, blades of grass, twine, paracord, ribbon, leather thong, spaghetti, ramen, copper wire, hairpin, herb stalk, fish bone, eel grass.\n\nLEFT surface: damp moss with twigs. RIGHT surface: polished slate plate. Each panel has its own directional lighting (slightly different angle / colour temperature). Per-object micro-variation in shade/sheen so corresponding curves never share a pixel-level signature. Background must not create new stroke-like shapes.\n\nNo text, no labels, no overlays.\n\n⚠️ LEFT-PANEL PRECISION (HIGHEST PRIORITY): The single template stroke on the LEFT must be reproduced with PIXEL-EXACT precision — same start point, same end point, same path, same length, same curvature, same thickness, same (x, y) position as the source stroke within ~2 px. Do NOT move it, do NOT recentre it, do NOT lengthen or shorten it, do NOT smooth or restyle the curve. The LEFT stroke is a precision reference — a curve overlay must align almost perfectly.", "Re-render as a photorealistic top-down photograph. Same vertical divider as source.\n\n⚠️ LOCKED CONSTRAINTS:\n • Scale: every rendered curve occupies the EXACT same pixel area as its source stroke. No rescaling, no normalising.\n • Divider: vertical separator stays at the source's exact x-coordinate. LEFT and RIGHT panel widths match the source.\n • Position: every curve stays at the same (x, y) and follows the same path/length/curvature within ~5 px.\n\nPANEL STRUCTURE:\n • LEFT (template): the ONE source stroke → ONE elongated object of class A on the LEFT surface.\n • RIGHT (field): every source stroke → ONE elongated object of class B on the RIGHT surface. Object count must equal source exactly — never add, skip, merge, or split.\n\nPick TWO different items from this pool — one for LEFT (class A), a different one for RIGHT (class B):\n yarn, embroidery thread, wool roving, cotton thread, silk thread, sinew, horsehair, fishing line, dental floss, jute twine, hemp string, linen thread.\n\nLEFT surface: tailor's wooden cutting mat. RIGHT surface: denim cotton sheet. Each panel has its own directional lighting (slightly different angle / colour temperature). Per-object micro-variation in shade/sheen so corresponding curves never share a pixel-level signature. Background must not create new stroke-like shapes.\n\nNo text, no labels, no overlays.\n\n⚠️ LEFT-PANEL PRECISION (HIGHEST PRIORITY): The single template stroke on the LEFT must be reproduced with PIXEL-EXACT precision — same start point, same end point, same path, same length, same curvature, same thickness, same (x, y) position as the source stroke within ~2 px. Do NOT move it, do NOT recentre it, do NOT lengthen or shorten it, do NOT smooth or restyle the curve. The LEFT stroke is a precision reference — a curve overlay must align almost perfectly.", "Re-render as a photorealistic top-down photograph. Same vertical divider as source.\n\n⚠️ LOCKED CONSTRAINTS:\n • Scale: every rendered curve occupies the EXACT same pixel area as its source stroke. No rescaling, no normalising.\n • Divider: vertical separator stays at the source's exact x-coordinate. LEFT and RIGHT panel widths match the source.\n • Position: every curve stays at the same (x, y) and follows the same path/length/curvature within ~5 px.\n\nPANEL STRUCTURE:\n • LEFT (template): the ONE source stroke → ONE elongated object of class A on the LEFT surface.\n • RIGHT (field): every source stroke → ONE elongated object of class B on the RIGHT surface. Object count must equal source exactly — never add, skip, merge, or split.\n\nPick TWO different items from this pool — one for LEFT (class A), a different one for RIGHT (class B):\n spaghetti, udon, soba, rice vermicelli, lasagna strips, churros, breadsticks, pretzel sticks, liquorice sticks, chow-mein noodles, fettuccine, linguine.\n\nLEFT surface: marble pastry slab. RIGHT surface: dark slate plate. Each panel has its own directional lighting (slightly different angle / colour temperature). Per-object micro-variation in shade/sheen so corresponding curves never share a pixel-level signature. Background must not create new stroke-like shapes.\n\nNo text, no labels, no overlays.\n\n⚠️ LEFT-PANEL PRECISION (HIGHEST PRIORITY): The single template stroke on the LEFT must be reproduced with PIXEL-EXACT precision — same start point, same end point, same path, same length, same curvature, same thickness, same (x, y) position as the source stroke within ~2 px. Do NOT move it, do NOT recentre it, do NOT lengthen or shorten it, do NOT smooth or restyle the curve. The LEFT stroke is a precision reference — a curve overlay must align almost perfectly.", "Re-render as a photorealistic top-down photograph. Same vertical divider as source.\n\n⚠️ LOCKED CONSTRAINTS:\n • Scale: every rendered curve occupies the EXACT same pixel area as its source stroke. No rescaling, no normalising.\n • Divider: vertical separator stays at the source's exact x-coordinate. LEFT and RIGHT panel widths match the source.\n • Position: every curve stays at the same (x, y) and follows the same path/length/curvature within ~5 px.\n\nPANEL STRUCTURE:\n • LEFT (template): the ONE source stroke → ONE elongated object of class A on the LEFT surface.\n • RIGHT (field): every source stroke → ONE elongated object of class B on the RIGHT surface. Object count must equal source exactly — never add, skip, merge, or split.\n\nPick TWO different items from this pool — one for LEFT (class A), a different one for RIGHT (class B):\n copper wire, aluminium wire, solder strips, steel cable, nylon zip-ties, fibre-optic strands, jumper wires, leather laces, ribbon cable, brass rod offcuts.\n\nLEFT surface: galvanised steel workbench. RIGHT surface: weathered concrete floor. Each panel has its own directional lighting (slightly different angle / colour temperature). Per-object micro-variation in shade/sheen so corresponding curves never share a pixel-level signature. Background must not create new stroke-like shapes.\n\nNo text, no labels, no overlays.", "Re-render as a photorealistic top-down photograph. Same vertical divider as source.\n\n⚠️ LOCKED CONSTRAINTS:\n • Scale: every rendered curve occupies the EXACT same pixel area as its source stroke. No rescaling, no normalising.\n • Divider: vertical separator stays at the source's exact x-coordinate. LEFT and RIGHT panel widths match the source.\n • Position: every curve stays at the same (x, y) and follows the same path/length/curvature within ~5 px.\n\nPANEL STRUCTURE:\n • LEFT (template): the ONE source stroke → ONE elongated object of class A on the LEFT surface.\n • RIGHT (field): every source stroke → ONE elongated object of class B on the RIGHT surface. Object count must equal source exactly — never add, skip, merge, or split.\n\nPick TWO different items from this pool — one for LEFT (class A), a different one for RIGHT (class B):\n twigs, dried pine needles, fern fronds, vine tendrils, leaf veins, herb stalks, dried wheat stalks, split bamboo strips, dried lavender stems, juniper sprigs.\n\nLEFT surface: damp dark forest soil. RIGHT surface: pale weathered birch plank. Each panel has its own directional lighting (slightly different angle / colour temperature). Per-object micro-variation in shade/sheen so corresponding curves never share a pixel-level signature. Background must not create new stroke-like shapes.\n\nNo text, no labels, no overlays." ] } } }