Spaces:
Running on Zero
Running on Zero
| # ZeroGPU only runs torch 2.8.0 and newer, which is why these pins differ from the repo's | |
| # pyproject.toml. diffusers stays at the paper's version: the SR LoRA is injected by matching | |
| # the class name "AdaLayerNormZero", so a diffusers rename would silently inject nothing. | |
| # `spaces` is deliberately absent. The platform pins its own copy and a second pin breaks pip. | |
| torch==2.8.0 | |
| torchvision==0.23.0 | |
| diffusers==0.32.2 | |
| transformers==4.49.0 | |
| accelerate | |
| peft>=0.13 | |
| safetensors | |
| huggingface-hub | |
| qwen-vl-utils | |
| # SD3's third text encoder is T5. Its fast tokenizer is built from a sentencepiece model, and | |
| # the conversion wants protobuf. Both arrive transitively in the research env, not here. | |
| sentencepiece | |
| protobuf | |
| lpips | |
| einops | |
| pyyaml | |
| tqdm | |
| numpy | |
| pillow | |
| imageio | |
| imageio-ffmpeg | |