Spaces:
Running on Zero
Running on Zero
| title: Wave Forcing | |
| emoji: 🌊 | |
| colorFrom: blue | |
| colorTo: red | |
| sdk: gradio | |
| sdk_version: 6.31.0 | |
| app_file: app.py | |
| pinned: false | |
| python_version: "3.12" | |
| startup_duration_timeout: 1h | |
| short_description: Streaming text-to-video with Wave Forcing on Wan2.1-T2V-1.3B | |
| license: apache-2.0 | |
| # Wave Forcing — Streaming Text-to-Video | |
| Demo of **Wave Forcing: Towards Speed-of-Light Streaming Video Generation** | |
| ([paper](https://huggingface.co/papers/2611.00859), [code](https://github.com/SJTU-DENG-Lab/WaveForcing)). | |
| Wave Forcing is a rolling-window distillation scheme that turns a block-causal video | |
| diffusion model into a streaming generator: each 3-frame block is denoised in only 4 or 5 | |
| steps while overlapping windows keep future blocks in-flight, so the model emits | |
| continuous video at high frame rates. This Space runs the authors' released | |
| [WaveForcing-1.3B checkpoints](https://huggingface.co/SJTU-DENG-Lab/WaveForcing-1.3B) | |
| (on the Wan2.1-T2V-1.3B backbone) through their vendored single-GPU reference | |
| pipeline (`naive_rt`), with both the 4-step and 5-step distilled variants. | |
| Enter a prompt, pick a distilled checkpoint and a length in latent frames | |
| (each latent frame is 4 pixel frames at 832×480, 16 fps), and generate. | |
| ## Notes | |
| - The paper's speed-of-light throughput numbers come from the authors' multi-GPU | |
| WaveRT runtime; this Space is the faithful single-GPU reference for the same | |
| weights, so latencies here are higher than the paper's benchmark numbers. | |
| - Example prompts are taken from the paper repo (Apache-2.0 license). | |
| - Code in `naive_rt/` is vendored from the | |
| [SJTU-DENG-Lab/WaveForcing](https://github.com/SJTU-DENG-Lab/WaveForcing) repository | |
| (Apache-2.0), trimmed to the single-GPU inference path. | |