[SANA-WM] Request: Release the 4-step distilled Stage-1 DiT checkpoint

Summary

The SANA-WM paper's "60s efficiency ablation" reports a large speedup from distilling the Stage-1 DiT from 60 steps down to 4 steps (DiT: 21.7 min → 43s, ~27x), which combined with the sink refiner and NVFP4 gives the advertised ~38x overall / RTX 5090 numbers.

However, the current open-source release (Efficient-Large-Model/SANA-WM_bidirectional) and the inference script (inference_video_scripts/inference_sana_wm.py) only ship the 60-step base Stage-1 DiT checkpoint (dit/sana_wm_1600m_720p.safetensors, --step default 60). There is no 4-step distilled Stage-1 checkpoint or config wired into inference.

Request

Could you please release the 4-step distilled Stage-1 DiT checkpoint (and its inference config / step schedule) used to produce the "+ distill" rows of the 60s efficiency ablation?

Why

• The docs state "the distilled variant runs on an RTX 5090 with NVFP4 quantisation," but no distilled checkpoint / inference path is currently available to reproduce this.
• Passing --step 4 to the existing 60-step base model only under-samples it rather than running the distilled model.

Thanks for the great work!