混元Video 1.5模型配置器

混元Video 1.5 is Tencent's 8.3B video model and the current spine of the HunyuanVideo family - the KB's take is that while Tencent keeps building on it, the local community has largely moved to Wan. But if your LoRA needs live on HunyuanVideo 1.5 (or its derivatives), this is the node that wires it into Diffusion-Pipe. It's the pack's most recently-updated video model node and shares its encoder family with HunyuanImage 2.1, so if you've set up one, the other is familiar.

What it does

Four required paths plus a shift value:

• diffusion_path - the main video model, e.g. hunyuan_video_1.5.safetensors.
• vae_path - the video VAE, e.g. hunyuan_video_1.5_vae_fp16.safetensors.
• Text_Encoder - and notice the example path: qwen_2.5_vl_7b.safetensors. Same Qwen2.5-VL encoder as HunyuanImage 2.1. If you already downloaded it for that node, you can reuse it here - one of the nicer side effects of Tencent standardizing on the same encoder across the family.
• ByT5 - the byte-level spellout model, e.g. byt5_small_glyphxl_fp16.safetensors, same file as Image 2.1.
• shift (default 1) - the timestep shift, and the tooltip is worth quoting because it's unusually opinionated for this pack: "a higher shift may improve video training results." That's the author's empirical hint, so if your first run looks soft, nudging shift up is a legitimate experiment before you touch the learning rate.

The output is the standard model_path config wired to GeneralConfig.model_config. The README table gives HunyuanVideo-1.5 the full set - LoRA ✅, full fine-tune ✅, fp8 ✅ - plus the changelog mentions Z-Image FP8 Comfy-format training is supported via the HunyuanVideo-1.5 model path.

Fitting it in

Video training needs the whole video toolkit, not just this node: FrameBucketsNode for clip lengths, video_clip_mode on GeneralConfig for how clips get extracted, and an EvalDatasetConfig with its own frame_buckets if you want a real curve. The chain stays the same - model node → GeneralConfig → Train.

Installing the pack

Shared install, Linux/WSL2 only, submodules mandatory:

cd ComfyUI/custom_nodes/
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
git submodule update
pip install -r Diffusion_pipe_in_ComfyUI/requirements.txt

Common issues

The pattern repeats: all four paths are required, full and absolute, and the node errors if any is empty - so read the console. The encoder mix-up is the subtle one: the Qwen2.5-VL 7B file is the same for Image 2.1 and Video 1.5, and grabbing a wrong Qwen variant gives you an opaque crash at load. And remember the pack's WSL2 drive-letter convention (Z:/..., not /mnt/z/...). If training starts but looks bad, the author's own advice is the shift knob - try it before you chase hyperparameters.