Updates for WAN 2.1 and Hunyuan image to video ComfyUI workflows.
The last two weeks have been pretty interesting with 3 great video models coming out: WAN 2.1, Hunyuan Image to Video and LTXV. Here are a few updates for the native workflows for two of these models.
Wan 2.1 in fp16
We noticed that these models work better when inferenced in fp16 versus bf16. This is a quality increase versus the original bf16 that we uploaded and gives results that are much closer to using the original fp32 weights with full precision compute. This is very apparent when using the small 1.3B model but you will also see visible improvements on the 14B models. If you are using the bf16 file I highly recommend upgrading to the fp16 one.
For those who don't have enough vram for the 16 bit weights we have uploaded some fp8_scaled weights which are a tiny bit slower than the regular fp8 weights but should give better results than the regular fp8 weights.
Look at the example page for updated workflows and how to download these new files: https://comfyanonymous.github.io/ComfyUI_examples/wan/
HunyuanVideo v2 version of their Image to video model
One day after initially released their long awaited image to video model the HunyuanVideo team released a new "fixed" version.
This new model which I call the v2 "replace" model is actually a new model with a slightly different architecture.
The v2 model seems to follow the guiding image more closely than their v1 model but the v1 model might have better motion and prompt following. Both are interesting and might be useful depending on what you want.
You can find workflows for both the original v1 "concat" model and the new v2 "replace" model on the examples page: https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/
Other News
If you think these models are too slow and missed the new LTXV model which came out last week workflows can be found: https://comfyanonymous.github.io/ComfyUI_examples/ltxv/
Also for those who want to save WebM files the core SaveWEBM node now shows the video on the node and the frontend can load the workflow from the saved webm.