mastaP/wan22-remix-sfw-t2v

Wan2.2 Remix (SFW) — Text to Video

A text-to-video demo for the SFW branch of `FX-FeiHou/wan2.2-Remix`, a community finetune of Wan 2.2 that targets human dynamics, realistic motion and scene consistency.

This Space loads only the repository's two SFW text-to-video checkpoints. The other weights published in that repository are not used here and are not reachable through this Space.

What it does

Give it a prompt describing a subject, a motion, a camera move and a lighting situation, and it renders a short clip at 832×480, 16 fps. Clip length is adjustable between 1 and 5 seconds.

How it runs

Wan 2.2 is a mixture-of-experts model: a high-noise expert handles the early denoising steps, a low-noise expert the later ones. Remix ships both as single-file ComfyUI checkpoints in fp8_e4m3, 14.29 B parameters each.

They are read as bfloat16 — which lets diffusers map the ComfyUI key layout onto its own config — and then re-quantized to float8 with torchao. Without that second step the two experts would need roughly 57 GB and overflow the 48 GB ZeroGPU allocation. VAE, text encoder, tokenizer and scheduler come from `Wan-AI/Wan2.2-T2V-A14B-Diffusers`.

The high-noise expert is the Lightning dyno variant, so low step counts are expected — 8 total steps at guidance 1.0 is the default.

AI-generated content

Every clip this Space produces is synthetically generated by an AI model. Under Art. 50 of the EU AI Act, such content has to be marked as artificially generated wherever it is published or shared. Please label it accordingly.

Do not use this Space to depict real, identifiable people without their consent.

Licensing

The base model, Wan 2.2, is Apache-2.0. The Remix finetune declares license: other and ships no license text, so the terms for its weights and their output are not established — treat the rights status as unresolved before using output commercially.

API and MCP

Every function is exposed over the Gradio API and as an MCP tool:

from gradio_client import Client

client = Client("mastaP/wan22-remix-sfw-t2v")
video, seed = client.predict("a lone hiker crests a ridge at golden hour", api_name="/generate_video")

Data processing

Prompts you enter are sent to this Space's backend on Hugging Face infrastructure and used solely to generate the requested clip. Nothing is stored beyond the lifetime of the request — Spaces are stateless and this one mounts no persistent storage. Hugging Face acts as the hosting provider and processes connection data (including your IP address) under its own privacy policy.

About this Space

A personal, non-commercial technical demo — not a product, not a service offering, and not operated on behalf of a business. Nothing here is sold, advertised or monetised.

Questions about the model itself belong in the model repository; questions about this Space go in its Community tab.