LTX-Video模型配置器

LTX-Video is the fast, video-gen model that actually fits on a consumer GPU, which makes it a favorite for people who want to train a video LoRA without renting a cluster. LTXVideoModelNode is the model picker that tells the pack "train on LTX-Video," and it's the node that decides between the two ways you can feed the model in: a full diffusers folder or a single safetensors file.

The pack's model support table marks LTX-Video as LoRA yes, full fine-tune no - so this node is aimed squarely at LoRA training. That's the honest framing: you're here to teach a fast video model a character or a style, not to re-train it from scratch.

How it works

Like every model node in this pack, it doesn't load any weights. It builds a model_path config dictionary - here {"type": "ltx-video", "diffusers_path": ...} - which you pass into a ModelConfig node, which merges in dtype and timestep settings and feeds GeneralConfig. The actual training weights get loaded later, by the deepspeed subprocess that the Train node launches.

The one field that's not pure plumbing is first_frame_conditioning_p. LTX-Video supports image-to-video (i2v) training, where the first frame conditions the generation. This value is the probability (0 to 1, default 1) that a given training clip uses its first frame as a conditioning signal. At 1.0, every sample is i2v-style; at 0, none are (pure text-to-video); anywhere between, you get a mix. For a character LoRA that should also work from a still image, some conditioning is usually a good idea - that's why the default is full-on.

The inputs that matter

• diffusers_path - the required one: full path to an LTX-Video diffusers model folder, e.g. /data/models/LTX-Video.
• use_single_file + single_file_path - flip this pair to instead point at a single safetensors like ltx-video-2b-v0.9.1.safetensors. Only one style at a time.
• first_frame_conditioning_p - the i2v mix described above.

Output: model_path → ModelConfig.

Install

Pack-wide - ComfyUI Manager (search "Diffusion_pipe_in_ComfyUI") or:

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

Linux/WSL2 only - this pack will not train on plain Windows. The heavy deepspeed/diffusers stack comes with it. If you want the community's take on where LTX fits, it's the low-VRAM video training lane - but expect to do your own dataset work, because that's where video LoRA quality actually lives.

Where people get burned

Two things. First, the use_single_file toggle: it only does something if you've also filled in single_file_path, and leaving both blank-but-toggled means the pack falls back to diffusers_path - which you also left blank, so you'll get a "diffusers_path cannot be empty" error. Fill one path style or the other. Second, if you're training i2v and your clips come out ignoring the reference frame entirely, check first_frame_conditioning_p - a value like 0.1 will mostly skip conditioning, which is what you'd want for a T2V-flavored LoRA, but not if you expected i2v behavior. Name the field's job: it's a probability, not a strength slider.