VideoLaVITI2VLong

VideoLaVITI2V gives you one 24-frame clip, which is about two seconds at a typical 12fps. Not exactly cinema. VideoLaVITI2VLong is the same idea stretched: it runs the image-to-video generation in multiple clips and splices them into one longer IMAGE batch. Same inputs as the short version, plus one new knob: clip_num.

How it works

The model is autoregressive over time - it was pre-trained on video as a sequence of tokens - so it can keep generating past the first clip by conditioning each new clip on what came before. The node calls multimodal_video_generate with clip_num and inverse_rate handling the continuity (the underlying code carries the previous clip's last frame forward, re-noised). Then it stitches: clip_num=2 gives you the first clip's 24 frames plus 23 more from the second, for 47 total. Notice the seam logic - it drops the first frame of each subsequent clip. That's not a bug in your workflow, it's literally what the wrapper does in the source.

The input that changes everything

• clip_num (default 2) - how many 24-frame chunks to generate. Each extra clip costs roughly the same generation time as a full short clip, so clip_num=3 is closer to 3× the runtime of the base node. There's no free lunch; "long video" here means "same slow generation, more times."

The rest are the familiar set: image, prompt, video_width/video_height (default 576×320, multiples of 64), guidance_scale_for_llm (4), num_inference_steps (50), top_k (50), seed (16). Output is IMAGE - all the stitched frames in one batch, ready for VHS_VideoCombine.

Install

Identical to the rest of the pack: ComfyUI Manager search "ComfyUI-LaVIT", or git clone https://github.com/chaojie/ComfyUI-LaVIT into custom_nodes, with the model in models/diffusers/Video-LaVIT-v1. Don't forget the manual pip install -r VideoLaVIT/requirements.txt - the pack's root requirements file is empty, so Manager won't install the pinned deps that make any of this run.

Where people get burned

Longer is not smoother. Because each clip is generated from the previous clip's tail, errors compound - a weird warp in clip one becomes a worse warp in clip three. The official repo is upfront that long-video generation is still being worked on. Practical advice: keep clip_num modest (2–3), keep motion in the prompt gentle, and expect drift over time. And a reminder that all the usual research-wrapper caveats apply - this is an unmaintained single-commit pack around a 2024 model trained on watermarked 320p video. It's a fascinating glimpse at token-based video generation; it is not your production video pipeline.