HunyuanVideo CLIP Loader

HunyuanVideo 1.5 is Tencent's late-2025 refresh of their open video model - a genuinely notable update, with a step-distilled variant that generates a 5-second 480p clip on a single 4090 in around 75 seconds, and native ComfyUI support. Like the rest of the current generation, it drives text conditioning with a language model rather than CLIP. For the text-to-video path this pack builds, that model is Qwen2.5-VL, and HunyuanVideoCLIPLoader is the node that loads it and presents it as a CLIP wire.

The node's own description carries the single most important warning up front: use Qwen2.5-VL, NOT Qwen3. They're both Qwen, they're both text encoders in this pack, and they are not interchangeable - Qwen3-4B is the Z-Image encoder, Qwen2.5-VL is HunyuanVideo's (and Qwen-Image-Edit's). Grab the wrong one and nothing downstream will work right.

How it works

It loads the Qwen2.5-VL weights and wraps them so ComfyUI's native HunyuanVideo sampler and VAE can consume the conditioning. There's also an optional byT5 slot: byT5 is a byte-level encoder that handles quoted text in your prompt (words you want rendered in the video) and adds multilingual coverage. It's optional because plenty of prompts don't need in-video text, and the node defaults it to None.

This is the entry point for the whole HunyuanVideo branch - load the encoder here, then everything else (HunyuanVideoTextEncoder, the prompt expander) hangs off the clip it produces.

The inputs and outputs that matter

• qwen_model (required) - the Qwen2.5-VL-7B-Instruct weights. The one field that matters, and the one people get wrong; make sure it's the VL model.
• byt5_model (optional, default None) - the byte-level encoder for in-video text and multilingual prompts. Leave it None unless you're rendering quoted text.

Output: clip - feed it into HunyuanVideoTextEncoder and, if you want prompt expansion, HunyuanVideoPromptExpander.

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. The models are the heavy part: you need the Qwen2.5-VL-7B-Instruct weights for the encoder, the HunyuanVideo 1.5 diffusion model and VAE (Comfy-Org publishes repackaged BF16/FP8 builds), and optionally a byT5 model. On a consumer card, the fp8 encoder and the step-distilled diffusion model are what make this runnable.

Common issues & troubleshooting

You loaded Qwen3 and it's broken. The number-one mistake, and the node warns about it explicitly. HunyuanVideo 1.5 (via this pack) wants Qwen2.5-VL. Qwen3-4B belongs to Z-Image. Swap it.

Quoted text isn't rendering in the video. That's the byT5 job. Load a byT5 model in the optional slot; without it, in-video text and some multilingual handling won't work as well.

The encoder dropdown is empty. ComfyUI only lists models it finds in the expected folder. Confirm the Qwen2.5-VL safetensors is placed correctly and restart to re-scan. And remember it's a 7B VLM - on tight VRAM, a quantized build is the difference between running and OOMing.