HunyuanVideo 1.5 Leo Vision Encoder Model Loader

HunyuanVideo 1.5 uses a SigLIP vision encoder to understand your reference image - the same encoder family behind Flux's image understanding - and HyVideo15VisionEncoderLoader is the node that loads it into the complete workflow. It's the "Leo Vision Encoder Model Loader" in the menu (a leftover display name from the wrapper lineage; it's this plugin's own loader). If you're doing image-to-video with the split nodes, this is where the model's eyes come from.

What it does

• path - folders under ComfyUI/models/clip_vision. None auto-downloads from black-forest-labs/FLUX.1-Redux-dev into models/clip_vision/hyvideo15/siglip. This is the one download in the pack that's gated: you need to accept FLUX.1-Redux-dev's license on HuggingFace and pass a token.
• vision_encoder_type (siglip) - siglip or None. The "None" option is for when you're pointing at a folder that already holds the encoder; siglip is the default and the only type auto-download supports.
• load_device (main_device / offload_device) - the siglip encoder is small compared to the Qwen text encoder, but offload_device is still a fine choice on tight VRAM.
• hf_token - your HuggingFace access token, required for the gated download to succeed for most people.

Output model (HYVID15VISIONENCODER) feeds HyVideo15VisionEncode, which produces the semantic vision states the transformer consumes. In the simplified HyVideo15ModelLoader, the same siglip encoder is loaded as part of the all-in-one config.

How it fits

HyVideo15VisionEncoderLoader → HyVideo15VisionEncode → HyVideo15Transformer. The reference image also flows through HyVideo15VaeEncode (pixel-level condition) and HyVideo15CFG, so the encoder here is only one of three places the image shows up - but it's the one doing the "semantic understanding," and the reason 1.5's I2V holds subject identity well across frames.

Installing it

Part of comfyui_hunyuanvideo_1.5_plugin. ComfyUI Manager → search "HunyuanVideo-1.5 nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yuanyuan-spec/comfyui_hunyuanvideo_1.5_plugin
cd comfyui_hunyuanvideo_1.5_plugin
pip install -r requirements.txt

Restart. Before first run, accept the FLUX.1-Redux-dev license on HuggingFace and have your token handy.

Common issues

This is the node where people hit the pack's most confusing failure: the auto-download shells out to the hf CLI and, if you haven't accepted the gated repo's license (or haven't set hf_token), it fails - often leaving a half-populated clip_vision/hyvideo15 folder that then blocks retries. The fix is to accept the license, set hf_token, delete the partial folder, and run again. The good news: it's a one-time pain. And remember the model-level context - HunyuanVideo 1.5 is a solid I2V model that the 2026 community mostly left for Wan/LTX, so this loader matters most if 1.5's first-frame fidelity is specifically what you want.