README.md

---
title: SGF+ Video Generation
emoji: 🎬
colorFrom: purple
colorTo: yellow
sdk: gradio
sdk_version: 6.29.1
app_file: app.py
python_version: "3.10"
startup_duration_timeout: 1h
short_description: Few-step autoregressive text-to-video with SGF+ dual experts
---

# SGF+: Decoupling Gradient Flows for Autoregressive Video Generation

Interactive demo of **SGF+** (chunkwise Self Gradient Forcing Plus), a few-step
autoregressive text-to-video model built on Wan2.1-T2V-1.3B. Each latent block
(3 latents = 12 pixel frames) is denoised in only **4 steps** by the generation
expert, then written back to the KV/cross-attention caches through a dedicated
**memory expert** — the paper's core contribution — before the block is decoded
and streamed to your browser as MPEG-TS chunks.

- 📄 [Paper](https://huggingface.co/papers/2610.10429)
- 💻 [Code (Apache-2.0)](https://github.com/Zihan-Su/Self_Gradient_Forcing_Plus)
- 🤗 [Model weights](https://huggingface.co/ZihanSu/Self_Gradient_Forcing_Plus)

## Usage

Type a prompt (or pick an example) and press **Generate video**. Frames start
streaming as soon as the first block is ready; 7 blocks produce an 81-frame
480×832 clip at 16 fps (~5 s of video). Use the **Blocks** slider for shorter
clips and the seed for reproducible generations.

The inference loop follows the authors' `inference.py` (chunkwise config
`sgf_plus_chunkwise.yaml`, EMA generator weights, warped 4-step denoising
schedule, per-block memory-expert KV refresh) — capped to the 21-latent clip
length the model was trained on, rather than the paper's long-horizon rollout.

## Credits

SGF+ code and prompts are from the [Self_Gradient_Forcing_Plus](https://github.com/Zihan-Su/Self_Gradient_Forcing_Plus)
repository (Apache-2.0). Base model: [Wan-AI/Wan2.1-T2V-1.3B](https://huggingface.co/Wan-AI/Wan2.1-T2V-1.3B).