HunyuanVideo 1.5 Model Loader

HyVideo15ModelLoader is the front door of this pack. It's the simplified loader that pulls in everything HunyuanVideo 1.5 needs in one go - the 8.3B diffusion transformer, the 7B Qwen2.5-VL text encoder, the siglip vision encoder, the 3D causal VAE, the flow-matching scheduler, and the byT5 glyph machinery - and hands you a single hunyuanvideo_model_config dict. Feed that into HyVideo15T2VSampler or HyVideo15I2VSampler and you're generating. It's also the node that does the auto-downloading this pack is known for: leave model_path empty and it fetches the entire model stack for you.

What it loads (and what to set)

• model_path - empty string means "auto-download everything to ComfyUI/models", which the README explicitly recommends. If you've manually placed the models, point this at your hyvideo-1.5 model folder. Note the first auto-download is large - we're talking tens of gigabytes across transformer, VAE, and encoders.
• resolution (480p) / task (t2v) - these pick the transformer subfolder (480p_i2v, 720p_t2v, etc.). For I2V set task to i2v; for the fast lane you can point model_path at a distilled variant folder.
• attn_mode (flash) - flash, ptm_sparse_attn (1.5's SSTA sparse attention, needs a supporting GPU), or flash3. Flash is the safe default; the pack README recommends installing FlashAttention.
• enable_offloading (false) - the big one for VRAM. This is what gets 1.5's official 14GB floor: turn it on and the loader stages models on CPU, shuffling them to the GPU as needed. enable_group_offloading adds block-group offloading of the transformer on top (and forces enable_offloading on).
• transformer_dtype (bfloat16) - bfloat16 is the right default; fp8-style savings aren't on the menu here, that's external quantized builds.
• text_encoder_type (qwen-2.5vl-7b) - this is the 7B Qwen2.5-VL MLLM, the model's "understanding" half. It's the biggest single memory consumer after the transformer. text_encoder_max_length (1000) and the hidden_state_skip_layer (2) defaults match the official recipe - leave them.
• vision_encoder_type (siglip) - the vision encoder from FLUX.1-Redux-dev, needed for I2V.
• hf_token - required if the FLUX.1-Redux-dev download needs the gated license accepted (it does, for most people).
• flow_shift (None) - auto-picks 5 at 480p, 7 at 720p. Leave it.

Output is hunyuanvideo_model_config (HUNYUANVIDEO_MODEL_CONFIG), which is the one input the samplers want.

The honest picture

This loader's convenience is real, but it also shows the pack's weight. Because the simplified model is all-or-nothing, you load the 7B text encoder even for T2V runs that could get away with less - which is exactly why the complete workflow splits the loaders apart. And on the model itself: HunyuanVideo 1.5 is a strong I2V model that's been outshone by Wan and LTX in the 2026 community, so consider this loader the fast route into a specific model, not the default video stack.

Installing it

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. The pinned deps (diffusers==0.35.0, transformers==4.57.1, torch>=2.6.0) need to land in ComfyUI's Python env.

Common issues

First-run OOM is the #1 complaint - flip enable_offloading on before blaming anything else. If the auto-download fails on the vision encoder specifically, that's the gated FLUX.1-Redux-dev repo: accept the license on HuggingFace and paste your token into hf_token. And a note on the license governing the models it fetches: the Tencent Hunyuan Community License excludes the EU, UK, and South Korea, and forbids training another AI model on outputs. Locally it's fine; read it before you commercialize.