Instructions to use Wan-AI/Wan2.2-TI2V-5B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Wan2.2
How to use Wan-AI/Wan2.2-TI2V-5B with Wan2.2:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Inference
- Notebooks
- Google Colab
- Kaggle
Wan 2.2 TI2V-5B running 100% client-side via WebGPU (C# / WebAssembly implementation)
Hey everyone! π
Following up on our LTX-Video release, I wanted to share our browser-based WebGPU implementation for Wan 2.2 TI2V-5B, built entirely in C# using SpawnDev.ILGPU.ML.
No Python environment, no PyTorch backends, and no server-side GPU processing. The 5B model weights stream directly into WebGPU compute buffers, letting users run video generation locally inside a browser tab.
π₯ Live Demo
You can test it right now in your browser (requires WebGPU support, e.g., Chrome/Edge):
π https://lostbeard.github.io/SpawnDev.ILGPU.ML/video
β‘ Supported Features in the WebGPU Studio:
- Text-to-Video & Image-to-Video: Generate or animate frames directly on-device using Wan 2.2 TI2V-5B.
- Turbo & MagCache Support: Runs accelerated workflows (e.g., Turbo 4-step modes or MagCache optimizations).
- Live Step Previews: Watch video latent frames form step-by-step as they denoise on your GPU.
- Client-Side Streaming: Weights stream dynamically over HTTP into browser storage (OPFS) and load straight into WebGPU buffers without touching the .NET heap.
π Performance
On a desktop GPU (e.g., RTX 4070):
- Wan 2.2 (640x352, 17 frames, 50 steps):
75 seconds (54 seconds with MagCache enabled).
βοΈ How it works
All matrix math and attention layers are compiled from C# down to WGSL kernels via SpawnDev.ILGPU.
The source code is fully open-source (MIT licensed) on GitHub:
π GitHub: https://github.com/LostBeard/SpawnDev.ILGPU.ML
Huge thanks to the Wan-AI team for releasing Wan 2.2! Would love to hear any thoughts or feedback on running larger video diffusion models directly inside WebGPU.
π
Really interesting to see the 5B model running in WebGPU. Thanks for documenting your preview workflow.
One possible way to plan longer or higher-resolution shots is to iterate at a lower resolution first, then compare local and rented-GPU options for the final render. Required memory depends on the model, precision, frame count and offloading; I would not infer a universal VRAM minimum from this preview. I have not tested this proposed workflow end to end.
Question for you: does the OPFS cache survive a browser update, or do users re-download the weights?
Disclosure: I work on CloudGPU (per-minute GPU rental).
Thanks for checking it out!
To answer your question: Yes, the OPFS cache completely survives browser updates.
Because OPFS is part of the standardized web storage specification, the browser treats it with the same persistence rules as IndexedDB. The 5B model weights remain securely cached locally and won't trigger a re-download unless the user explicitly clears their browser cookies/site data, or if the device runs into an extreme, global low-storage emergency.
We highly recommend testing it out. It's a great showcase for how client-side WebGPU can drastically offset and even eliminate traditional per-minute cloud rendering costs! π