HunyuanVideo 1.5 Leo Text Encoder Model Loader

HunyuanVideo 1.5 doesn't use a small CLIP-style text encoder. Its "understanding" half is a full 7-billion-parameter MLLM - Qwen2.5-VL-7B-Instruct - and HyVideo15TextEncoderLoader is the node that loads it into the complete workflow. This is the "Leo Text Encoder Model Loader" in the menu, a leftover label from the wrapper lineage this pack descends from. It's the heavyweight of the loaders: roughly 15GB of weights, which is why 1.5's memory story is what it is.

What it does

• path - a dropdown of folders under ComfyUI/models/text_encoders. Set to None and it auto-downloads Qwen/Qwen2.5-VL-7B-Instruct into models/text_encoders/hyvideo15/llm. Set to a real folder to use a manual download.
• text_encoder_type - llm or None. "None" is the "I'm pointing you at a folder that already contains the encoder" option; llm is the default. Leave it.
• load_device - main_device or offload_device. Offload keeps the 7B model on CPU/RAM and stages it only when needed; choose it if VRAM is your binding constraint.

Outputs are text_encoder and text_encoder_2 - and the second one is always None. That's not a bug, it's a compatibility slot: the original Hunyuan wrapper concept allowed a second encoder, this pack just doesn't use one. HyVideo15TextEncode accepts both ports so the graph stays uniform. Wire text_encoder_2 anyway to keep the workflow tidy, or leave it dangling; it does nothing.

How it fits

HyVideo15TextEncoderLoader → HyVideo15TextEncode → HyVideo15Transformer. In the simplified HyVideo15ModelLoader, this same Qwen2.5-VL encoder gets loaded as part of the all-in-one config - one of the reasons that loader is so heavy. The split version exists so you can, for example, offload the encoder while the transformer stays resident, or swap in a different folder layout.

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, and give the first auto-download time - 15GB doesn't move fast on most connections. The README suggests HF_ENDPOINT=https://hf-mirror.com for slow regions.

Common issues

The big one is disk, not code: this download is ~15GB and the loader checks folder existence, not completeness, so an interrupted download leaves a partial folder that blocks future auto-downloads with a confusing transformers error downstream. If that happens, delete models/text_encoders/hyvideo15 and re-run. The other real decision is load_device: with offload off, this node alone can blow a 12GB card before the transformer even loads. Turn offload on unless you have headroom to burn. And note the bigger context - this 7B encoder is exactly why 1.5's "14GB floor" is a floor and not a recommendation: it's a serious model that needs serious hardware, more so than the Wan/LTX alternatives most people landed on.