Spaces:
Running on Zero
Bring back the early pro realism ai.
Seconded
Thirded
Thanks for the report — I reproduced and fixed both symptoms.
There were two separate causes:
Accelerator failure: the unquantized BF16 pipeline required ZeroGPU to transfer about 57.7 GB of tensors. The task could be rejected before
infer()even started. The transformer and text encoder are now dynamically quantized to FP8 before ZeroGPU binding, and the cold-call duration ceiling was raised. The model stack is pinned so it cannot silently switch behavior again.Psychedelic/grid output: I reproduced this with a tiny 61×68 source image and forced low output dimensions. With a proper 1024×704 source and automatic sizing, the same pinned model produced a clean 1248×864 edit while preserving the subject, pose, lighting, and background. The width/height controls have therefore been removed, sizing is automatic, and inputs smaller than 256 pixels on either side are rejected with a clear message.
The final deployed revision is 3c1179c. Live verification on ZeroGPU now shows:
- proper-resolution edit succeeds in about 14 seconds
- thumbnail and missing-image requests fail cleanly before GPU inference
- no unhandled runtime exception
- Space remains
RUNNING
Please retry with an input at least 256 px on each side. If a normal-resolution image still produces a grid, share the input dimensions, inference settings, and a non-sensitive description of the edit so I can reproduce that exact path.
The psychedelic grid is still there And now it also has oil painting filter.Screenshot_2026-07-31-15-35-40-833_com.android.chrome
