VideoLaVITI2I

Let's be blunt about VideoLaVITI2I: the name is wrong and the code is broken. Despite the "I2I" in its name, it has no image input - it calls the model's generate_image function, which is a pure text-to-image path. And worse, the node as shipped crashes when you run it. It's the one node in this pack you should treat as a curiosity, not a tool.

Why it breaks

Look at the source: VideoLaVITI2I.run accepts model, prompt, width, height, guidance_scale_for_llm, num_inference_steps, top_k, temperature, seed - then calls model.generate_image(...) passing a guidance_scale_for_decoder= argument that was never defined in that scope. That's a NameError the moment the node executes. It also takes the temperature input and ignores it, hardcoding temperature=1.0 instead. The author's single commit from April 2024 shipped it this way and never fixed it.

So your options: skip it, or patch it yourself. If you want it working, you can add a guidance_scale_for_decoder input (default 3.0 - that's the generate_image default) and pass the actual temperature through. It's a two-line fix in nodes.py and a restart. That's a real statement about this pack's quality: the headline generation nodes work, but this one is a field test.

What it's supposed to be

A text-to-image still at 1024×576 - the keyframe-only half of the T2V pipeline, without the motion. In Video-LaVIT's design, an image is just a keyframe with no motion tokens, so the model genuinely can do T2I. If you patch it, it's a perfectly usable still generator with that distinctive "token-LLM then decode" look. temperature would control token sampling diversity for the visual tokens. But even working, it's the least useful node here: the pack's whole pitch is video, and there are far better T2I models from 2024, let alone now.

The inputs

• model - a VideoLaVIT from VideoLaVITLoader.
• prompt - default "FPV drone footage of an ancient city in autumn," which is a video-flavored prompt for a still image. Change it.
• width / height - 1024×576 default, multiples of 64.
• guidance_scale_for_llm (4), num_inference_steps (50), top_k (50), temperature (1.0), seed (16).

Output is IMAGE - a single frame, saveable or wireable into I2V as a starting point.

Install

Pack install is unchanged: ComfyUI Manager search "ComfyUI-LaVIT", or git clone https://github.com/chaojie/ComfyUI-LaVIT into custom_nodes, model in models/diffusers/Video-LaVIT-v1, and manual pip install -r VideoLaVIT/requirements.txt because the root requirements file is empty.

Where people get burned

By the node itself. Anyone who wires an image into it will find the image input doesn't exist; anyone who runs it as-is hits the NameError. If a shared workflow uses it and you're staring at a red node, now you know why. The honest recommendation: use VideoLaVITT2V's keyframe path or any real T2I model instead, unless you specifically want to fix a research wrapper for the fun of it - which, fair, can be a fun afternoon.