# HiDream-O1-Image-Dev (Q8 MLX) — evaluation

**Setup:** lab branch `perf-lab-hidream-o1-mlx`, mlx-vlm 0.5.0 + mlx 0.31.2, Mac Studio (64 GB).
**Recipe:** Dev — 28 steps, FlashFlowMatch, `s_noise=7.5`, `noise_clip_std=2.5`, `shift=1.0`.
**All times** are honest wall-clock with `mx.eval` per step. **All RAM** is peak `maximum resident set size`.

## Q6 showcase verification (2026-05-09 evening)

Re-ran the same 10-prompt battery at Q6 with identical seeds. **All 10 are visually equivalent or better than the Q8 versions:**

- 9/10 are near-pixel-identical aesthetics (different latent noise from quant differences yields same compositions / lighting / subjects)
- **10 (text rendering) is visibly better at Q6** — "BLOOM CAFE" neon sign is crisp at Q6 vs a glitched "M" at Q8

Per-image timing was rock-steady at **35.9 s** (1.28 s/step). Total battery time: ~6 minutes vs ~12 minutes at Q8.

Outputs: `sample_outputs/showcase_q6/` (compare against `sample_outputs/showcase/` for the Q8 originals).

## Battery: 10 prompts, 1024×1024, all Q8

| # | Genre | Prompt summary | Result | Time |
|---|---|---|---|---|
| 01 | photo portrait | elderly Japanese tea master | **Excellent** — face character, gentle smile, paper screens, calligraphy | 81.5 s* |
| 02 | anime / illustration | pink-haired girl on Tokyo rooftop at dusk | **Excellent** — anime style + cherry blossoms + neon city below | 65.3 s |
| 03 | macro photo | dewdrop on spiderweb | **Excellent** — refractions, blurred leaf bg, crisp web detail | 65.9 s |
| 04 | architecture | futuristic library, holographic displays | **Excellent** — vaulted ceiling, stained glass, holo screens | 66.3 s |
| 05 | surreal painting | whale floating over desert at sunset | **Excellent** — magical realism, painterly clouds | 65.8 s |
| 06 | food flatlay | rustic Italian breakfast on marble | **Excellent** — golden croissants, espresso, berries, soft light | 66.4 s |
| 07 | cinematic action | samurai mid-leap with katana, Mt. Fuji bg | **Excellent** — dynamic pose, cherry blossoms, real mountain | 66.1 s |
| 08 | fantasy | dragon on crystal mountain with aurora | **Excellent** — iridescent scales, snow swirling, aurora visible | 66.4 s |
| 09 | wildlife photo | snow leopard staring at camera | **Excellent** — direct gaze, falling snow, mountain bg | 67.1 s |
| 10 | text rendering | "BLOOM CAFE" pink neon diner | **Good** — sign legible (small "M" glitch), retro diner, rainy street | 67.1 s |

*Image 01 included cold model load (~12-15 s).

**Steady-state per-image: 65-67 s at 1024×1024 Q8.** Dead-consistent across genres.

## Honest timings

| Resolution | Quant | Per step | Total (28 steps) | Peak RAM |
|---|---|---|---|---|
| 512×512 | Q4 | 0.89 s | 24.9 s | ~6 GB |
| 1024×1024 | Q4 | 2.37 s | 66 s | ~6 GB |
| 1024×1024 | **Q6** | **1.30 s** | **36 s** | **~8.5 GB** |
| 1024×1024 | Q8 | 2.36 s | 66 s | ~11.5 GB |
| 1280×704 | Q8 | 2.53 s | 70.7 s | ~7 GB |
| 704×1280 | Q8 | 2.35 s | 65.9 s | ~3 GB (warm cache) |
| 2048×2048 | Q4 | 8.44 s | 236 s | ~7.2 GB |
| 2048×2048 | Q8 | 9.86 s | 276 s | ~10.8 GB |

**Q6 is the sweet spot.** 2× faster than Q8 at 1024 with the same prompt fidelity (cat in sunlit kitchen + beach with palm trees both rendered identically to Q8 outputs). 30% less RAM. The bandwidth-bound theory holds: fewer bits per param → less weight bandwidth → faster per-step.

**Q4 corrupts brightness** (ships dark) so the speed of Q4 vs Q6 is academic — never use Q4 for production. Q6 has the speed and Q8 has the steady-state safety; Q6 wins on perf, Q8 wins on a deterministic upper bound on RAM.

## Where HiDream-O1-Image-Dev shines

- **Subject identity** — every prompt subject was rendered correctly. No "vibrant orange tabby" → cat-shape-blob. The model knows what things look like.
- **Multi-element scenes** — samurai + Fuji + cherry blossoms; cyberpunk Alice + neon Cheshire cat + circuit dress + rain. Composition stays coherent.
- **Style adherence** — anime ≠ photorealism ≠ oil painting ≠ macro. Got all four right.
- **Light realism** — the architecture image's light through stained glass; the food flatlay's morning warmth; the action scene's sunset rim lighting. Light feels real, not stamped on.
- **Text rendering** (limited) — "BLOOM CAFE" in neon was readable. Better than most diffusion models; not as clean as a model with explicit OCR pretraining.

## Where it's weak

- **Patch-grid artifact** in flat regions. PATCH_SIZE=32 with no overlap → visible 32×32 grid in skies, water, walls. Most visible at low-frequency content. Architectural — not fixable without retraining or an overlap-blending postprocess.
- **Q4 brightness collapse** — Q4 desaturates and darkens everything. Q8 fixes it. **Ship Q8.**
- **Hands** — hands when present in scenes (e.g. tea master holding cup) look fine at moderate detail, but the model isn't immune to the standard diffusion hand failure modes; haven't stress-tested.
- **Dense long text** — "BLOOM CAFE" is short and structured. A paragraph of text would likely fall apart.
- **Speed at 2048** — 4 minutes per image is slow for iterative work. Fine for a final pass.

## Sweet spot

**1024×1024, Q6, default Dev recipe, ~36 s/image, ~8.5 GB RAM.** Bright/colourful output equivalent to Q8, half the wall time, 30% less RAM. 512 is fast (~25 s) but loses detail. 2048 is gorgeous but iterative-unfriendly.

**Quant decision tree:**
- 16 GB Mac → don't run HiDream; use mflux Z-Image-Turbo
- 32 GB Mac → Q6 is comfortable, Q8 leaves no headroom alongside LTX
- 64 GB Mac → Q6 default; Q8 only when you want deterministic upper-bound RAM

## A/B vs mflux Z-Image-Turbo

Same prompts, same seeds, both at 1024×1024.

| # | Prompt | HiDream Q8 | Z-Image-Turbo Q4 (mflux) | Subjective winner |
|---|---|---|---|---|
| 1 | tea master | [v3](../sample_outputs/showcase/01_portrait_photo.png) — wide scene, paper screens, calligraphy | [zimg](../sample_outputs/ab_mflux/01_portrait_zimage.png) — tighter portrait, gray garment, smile | **Tie** — different framings, both excellent |
| 2 | sunlit beach | [v3](../sample_outputs/v3_1024_beach_q8.png) — turquoise water, palm trees, beach chair | [zimg](../sample_outputs/ab_mflux/02_beach_zimage.png) — vivid blue water, palms, big sand foreground | **Tie** — both nail the prompt |
| 3 | alice cyberpunk | [v3](../sample_outputs/v3_alice_horizontal_q8.png) (horizontal) — clear dress + face + Cheshire | [zimg](../sample_outputs/ab_mflux/03_alice_zimage.png) — more painterly, atmospheric Cheshire silhouette | **HiDream** for face/dress detail; **Z-Image** for atmosphere |

### Speed + RAM (measured, not estimated)

| Engine | Steps | Wall (1024) | Per step | Peak RAM |
|---|---|---|---|---|
| HiDream-O1-Dev / Q8 | 28 | **67 s** | 2.41 s | **11.5 GB** |
| Z-Image-Turbo / Q4 | 9 | 80 s | 8.85 s | **5.9–29.4 GB** (varies by prompt) |

Surprises:
- HiDream is **faster per image** despite needing 28 steps vs Z-Image-Turbo's 9 — Z-Image's per-step cost is ~3.7× HiDream's.
- Z-Image's peak RAM **varied wildly across prompts** (5.9 GB for portrait, 29.4 GB for the alice cyberpunk). HiDream's peak was steady at ~11.5 GB regardless of prompt complexity.

### Verdict

Both are excellent local engines. Pick by the workload:

- **Default/compact**: keep **Z-Image-Turbo** — 5.9 GB RAM on most prompts, runs anywhere.
- **Hero shots / max prompt fidelity**: **HiDream-O1-Q8** — faster wall time, deterministic memory, more environmental detail in the output.
- **Editing / multi-ref**: keep **mflux qwen-edit** — HiDream lab pipeline doesn't support refs yet.

## Patch-grid post-blend experiment

Implemented `--blend-seams <radius>` post-process in `generate_hidream_o1_mlx.py`: after decoding the final image, average a thin band across each 32-pixel patch boundary line (radius=1 → blend the seam row with one neighbour on each side, then 50% blend back into the seam itself).

**Result on the same beach prompt + seed 11 + Q8:**

| Comparison | Mean abs diff (out of 255) |
|---|---|
| baseline vs blend r=1 | 0.18 |
| baseline vs blend r=2 | 0.23 |

Per-row breakdown confirms the blend is **surgical** — only seam rows (every 32) change, by 1–2.7 pixel values; non-seam rows shift by <0.2. So the math is doing exactly what it says.

**But visually**: at Q8 the seam artifact is already mild. The blend's 1–2 pixel-value smoothing is below visual threshold. No win, but no harm — and zero added latency (numpy vector ops on a 1024×1024 image are sub-ms).

Bottom line: kept as opt-in flag `--blend-seams 1`. Did not enable by default. The real fix for the patch grid would need overlap-blended patches (architectural change) or a stronger spatial filter (which would visibly blur the image).

## Software-side speed: nothing left

Tested `mx.compile` on the forward pass: **0% improvement** (2.366 s/step compiled vs 2.368 s/step uncompiled). The forward is already bandwidth-bound by the 36-layer Q8 decoder's matmul stream — MLX is already at near-GPU-saturation. Same conclusion for `mx.fast.scaled_dot_product_attention` (already used inside mlx-vlm's Qwen3VLAttention).

**The path to faster is architectural, not algorithmic:**
- Fewer steps (would need a smaller distillation; Dev is already the distilled variant)
- Smaller backbone (would need re-distillation onto a 4B Qwen3-VL — no public version)
- Caching the text-portion hidden states across denoising steps — possible but invasive (would need to subclass mlx-vlm's Qwen3VLModel; ~2-5% speedup at best since text is <2% of seq length)

## Verdict

- **Working.** Q8 produces real, prompt-faithful, high-quality images at ~67 s/1024.
- **No more easy speedups.** The lab's inference loop is already at the floor for this architecture on this hardware.
- **Patch artifacts are real but mild.** Low-frequency regions show a 32-pixel grid. Subjects-with-content scenes hide it well.
- **Q8 is the only acceptable quant.** Q4 ships dark. If we ever want a smaller variant, would need different bit packing or selective Q6.

## Recommendation for Phosphene

Slot it in as a third local engine alongside `mflux Z-Image-Turbo` (compact tier) and `mflux FLUX.2-klein-4B` (comfortable tier). Mark HiDream as **comfortable+** (32 GB+) due to the 11.5 GB working set. Don't make it the default — it's slower per image and uses more RAM than Z-Image-Turbo. Make it **the option** for users who want max prompt fidelity and license clarity (MIT, no NC restriction).

See [PHOSPHENE_INTEGRATION_PLAN.md](PHOSPHENE_INTEGRATION_PLAN.md) for the patch.