Wan-S2V is an AI video generation model that can transform static images and audio into high-quality videos.
WIP: working on description adding all needed infos/tools! Use with some caution 🤪
Note: S2V has a very high chance of producing some 1st "flashy" over-saturated frames. That seems a limitation of all Wan 2.2 S2V models right now.
Requirements:
lite lorafor 4/8-step operation (optional)Main Model Wan2.2-S2V-14B
ComfyUI/models/unetGGUFAudio Encoder wav2vec2_large_english
ComfyUI/models/audio_encodersEncoder Umt5-xxl
ComfyUI/models/text_encodersWan2.1_VAE.safetensors
ComfyUI/models/vae
Usage hints:
Audio file should be about same length as the video file in seconds
👂🎶 👉 Hint: Click the sample for full-screen and play from the post with SOUND ON!
Sources:
Clip: https://huggingface.co/city96/umt5-xxl-encoder-gguf/
Model: https://huggingface.co/QuantStack/Wan2.2-S2V-14B-GGUF/
Lite LoRA: https://huggingface.co/calcuis/wan2-gguf/
YOU are responsible for outputs as always! If you make ToS violating content and I get aware I WILL report this.
Description
FAQ
Comments (1)
Remember, with Blackwell (and maybe earlier architectures) do NOT use scaled FP8 models- these massively downgrade prompt adherence and lip movement. Even a smaller GGUF Clip or main model will work much better- even with a gofaster LoRA.
This workflow does uses GGUF for both, in the mistaken understanding that small models are needed for 'low' VRAM like 12-16GB - no, so long as you have enough system RAM to cache the model, the slow iteration times (many times greater than 1 second) allow the model to stream as needed with no impact on iteration speed!
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.