HunyuanVideo Vision Loader (SigLIP)

If you want HunyuanVideo 1.5 to animate a starting image rather than generate from text alone, the model needs to actually see that image - and for that it uses a vision encoder. HunyuanVideoVisionLoader loads the SigLIP vision model and hands it out on a CLIP_VISION wire. It's the image-to-video half of the loaders; the text encoder (HunyuanVideoCLIPLoader) handles your words, this one handles your reference frame.

SigLIP is a sigmoid-loss vision encoder - think of it as the "eyes" that turn your input image into features the video model can condition on. Image-to-video is the workflow most people actually want for consistent results, because anchoring the first frame to a known image solves a lot of the drift you get from pure text-to-video, so this loader shows up whenever I2V is the goal.

How it works

It reads the SigLIP weights and exposes them as CLIP_VISION, the standard ComfyUI type for image conditioning encoders. Downstream, the HunyuanVideo I2V path uses those image features alongside your text conditioning so the generated motion starts from - and stays anchored to - your reference frame. There's exactly one thing to set here, which is correct for a loader: pick the model file and go.

The inputs and outputs that matter

• vision_model (required) - the SigLIP vision encoder file (the pack expects a patch14-384 SigLIP build, e.g. sigclip_vision_patch14_384.safetensors). Put it in your clip-vision models folder and select it.

Output: clip_vision - a CLIP_VISION wire for the HunyuanVideo image-to-video conditioning path.

How to install it

ComfyUI Manager: search ComfyUI-QwenImageWanBridge, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/fblissjr/ComfyUI-QwenImageWanBridge

then restart. You need the SigLIP vision model on disk (in the clip-vision folder), on top of the usual HunyuanVideo 1.5 stack - Qwen2.5-VL text encoder, the diffusion model, and the VAE. The vision encoder is small compared to the 7B text encoder, so it's not the thing that'll break your VRAM budget.

Common issues & troubleshooting

The dropdown is empty. The SigLIP file needs to be in the folder ComfyUI scans for clip-vision models. Drop it there and restart.

You only want text-to-video. Then you don't need this node at all - the SigLIP encoder is specifically for the I2V path. Pure T2V runs off the text encoder alone.

I2V ignores your starting image. Make sure the clip_vision output is actually wired into the image-to-video branch of your HunyuanVideo graph, and that you're feeding a reference image into that branch. A loaded-but-unconnected vision encoder does nothing; the image conditioning has to reach the sampler for the first frame to anchor.