akhaliq/Cosmos3-Nano

Add imageio and imageio-ffmpeg to requirements.txt

Optimize for ZeroGPU: dynamic duration, practical defaults (960x544, 121 frames, 25 steps)

Reduce inference steps to 30 (UniPC scheduler is designed for low step counts)

Align defaults with diffusers docs: 1280x720 HD, 189 frames, remove guidance_scale (use pipeline default)

Fix: pass num_inference_steps to text_to_image endpoint (was falling back to pipeline default of 35)

Set default inference steps to 100 and use HuggingFace avatar as logo

Speed up: 25 steps, 85 frames, 640x384 defaults

Add av (PyAV) for encode_video audio muxing

Use xlarge ZeroGPU (96GB) for Cosmos3 model

Fix: import spaces before torch to avoid CUDA init conflict

Add Cosmos3-Nano Gradio Server app with cinematic frontend

initial commit