Default prompt: 'detailed fur' (drop 'ultra-')

Set PYTORCH_CUDA_ALLOC_CONF=expandable_segments to fight fragmentation; explicit low_cpu_mem_usage=False on TAEF1 + add accelerate to requirements to silence the warning

Default prompt: cat now wears a woolen hat

Per-resolution GPU size: 512 → large (1× quota), 1024 → xlarge (2× quota); plain dispatcher routes between two @spaces.GPU entrypoints

Resolution as 512/1024 radio (no intermediates); clarify in info that PiD picks 2k vs 2kto4k accordingly

Preload both PiD checkpoints (2k + 2kto4k); pick at request time based on resolution (>512 → 2kto4k for 4K target)

Reduce duration to 90s (fits comfortably on full Blackwell)

Use size='xlarge' (full 96GB Blackwell instead of 48GB MIG; 2× quota cost)

Drop VAE tiling (it fits at normal decode); PiD has no tiling API to apply during upscale

Drop VAE tiling cleanup (mutation is fork-local on ZeroGPU; parent state untouched)

Enable VAE tiling+slicing only when resolution >= 768 (inside @spaces.GPU); reset after the call

Initial preview label: 'Z-Image with PiD' (only runtime updates say 'Generating…')

Drop prompt label; cap app width at 1200px

Shorter description; add 'Randomize seed' checkbox that writes the picked seed back into the Seed field

Layout: prompt + run inline above image; advanced settings in accordion below; match slider max_height to height

Bump preview + slider heights to 720/900px

Stream PiD's 4 internal student-sampler steps (yield after each); 'Upscaling with PiD — step K/4' label

Drop intermediate captures + gallery; use gr.ImageSlider for Z-Image vs PiD A/B; dynamic 'Generating Z-Image step X/N' / 'Upscaling' labels

Stream taef1 previews from Z-Image (thread+queue) into a flash gr.Image; swap to gr.Gallery only when all PiD steps done

Fix _latent_to_pil squeeze dim (T=1, dim=1 not dim=0); stream gallery as each PiD step completes

Patch Gemma2.forward to put torch.tensor on the embedding's device (ZeroGPU rejects the CPU normalizer cross-device multiply)

Replace transformers' vmap-based mask broadcaster with explicit broadcasting (ZeroGPU __torch_function__ can't fake-alloc inside vmap)

Force eager attention on Z-Image text encoder (vmap conflict with ZeroGPU); add Citrus theme + dark-mode color css

Initial PiD + Z-Image step-by-step denoising demo for ZeroGPU