README.md

metadata

title: DistillAlign — Autoregressive Video Distillation
emoji: 🎬
colorFrom: green
colorTo: pink
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Few-step causal text-to-video with a 1.3B student
python_version: '3.12'
startup_duration_timeout: 1h

🎬 DistillAlign — Autoregressive Video Distillation

Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation

This Space runs the released DistillAlign generator: a Wan2.1-T2V-1.3B student distilled for few-step causal (autoregressive) text-to-video generation. It samples 21 latent frames in 7 causal blocks with a 4-step DMD schedule (1000 → 750 → 500 → 250), keeps a KV cache across blocks, and streams the latents through the causal Wan VAE — 81 frames / 480×832 / 16 fps ≈ 5 s of video.

Links

• Model: LiJiaxing/DistillAlign
• Code: LiJiaxing0213/DistillAlign
• Paper: arXiv 2607.26811 · project page

Notes

• Prompt style: the model was distilled on long, detailed captions. Short prompts work but detailed multi-sentence descriptions give markedly better results — the example prompts below are the authors' own evaluation prompts.
• Speed: generation takes roughly a minute on the ZeroGPU RTX PRO 6000 (7 blocks × 4 denoising steps, plus streaming VAE decode).
• Seed: set a seed to reproduce a video, or leave 0 for a random one.
• Checkpoint used: checkpoints/distillalign_distill_1p3b_teacher.pt — the final joint-distilled 1.3B generator (1.3B DMD teacher), loaded with strict=True into a causal Wan2.1-1.3B wrapper.

Credits & licenses

This is a demo wrapper around the authors' code. The example prompts are the authors' own evaluation prompts from prompts/distribution_eval_16.txt in the official repository (Apache-2.0). The DistillAlign code and released weights are Apache-2.0; upstream Wan2.1-T2V-1.3B weights remain subject to their own license. The causal VAE decoder wrapper follows the Self-Forcing (Apache-2.0) implementation.