Add torchvision (required by WAN repo)

Stream SPEED output ahead of fullres via generator yields

load_pipeline: accept plain "cuda" device string for ZeroGPU

Bump Gradio to 5.34 for gradio_client schema fix (bool additionalProperties)

Restore huggingface_hub <1.0 pin (Gradio 5 still uses HfFolder)

Bump to Gradio 5 (schema bugfix + drops pydub/audioop dep)

Stub FLUX_DIR/WAN_CKPT/WAN_PATH env vars so load_config does not error on Spaces

Pin huggingface_hub <1.0 for gradio 4.x HfFolder compat

Pin python 3.11, audioop-lts fallback, update README title

Initial SPEED Gradio demo (FLUX + WAN 2.1, always-compare)

initial commit