VideoLaVITVideoDetokenizerLoader

In Video-LaVIT's architecture, video gets compressed into tokens: a keyframe plus motion tokens that describe how it moves. Something has to turn those tokens back into actual pixels - that's the video detokenizer, a diffusion model trained for exactly that job. VideoLaVITVideoDetokenizerLoader is the node that loads it. Its only real output is a VideoLaVITVideoDetokenizer object that feeds exactly one downstream node: VideoLaVITVideoReconstruction.

How it works

The loader points at ComfyUI/models/diffusers/<model_path> - the same model folder the generation loader uses - and specifically grabs the video_3d_unet.bin file inside it. That's the released detokenizer weight, a 3D UNet that conditions on a keyframe plus motion vectors and diffuses out a sequence of frames. It's loaded in fp16 and moved to CUDA. Nothing fancier: it's a dedicated model file, not the full Video-LaVIT bundle, and it exists because the research team released the tokenizer and detokenizer as separate, reusable components.

There's exactly one input: model_path, the same dropdown of folders in models/diffusers you see on every loader in this pack, defaulting to Video-LaVIT-v1. If the video_3d_unet.bin file isn't in that folder, the loader fails - the path is hardcoded to that filename.

Why you'd reach for it

By itself, nothing happens. This node exists to unlock the reconstruction workflow - the pack's most interesting demo, where you feed a real video in and the model re-renders it from one keyframe plus motion vectors. That's a compression story, not a generation story: it shows how well the tokenizer captured the video, and it's genuinely cool to watch the model rebuild a clip from a single frame and some motion data. If you're not running VideoLaVITVideoReconstruction, you don't need this node.

Install

Same pack install as everything else: ComfyUI Manager search "ComfyUI-LaVIT", or git clone https://github.com/chaojie/ComfyUI-LaVIT into custom_nodes. The full model download includes video_3d_unet.bin, but the reconstruction workflow has one extra dependency the other workflows don't: it needs the motion-vector-extractor package and an FFmpeg build that can re-encode video with the MPEG-4 Part 2 codec, because motion vectors are extracted by decoding the compressed bitstream. Those come from VideoLaVIT/requirements.txt plus your system FFmpeg, not from the pack's empty root requirements file.

Where people get burned

The missing video_3d_unet.bin is the usual failure - if you download only part of the model repo, or a snapshot predates the detokenizer release, the loader errors on the file path and it looks like a dependency issue. Re-run the full download. And if you land here expecting a "video enhancer" or an upscaler, you're in the wrong place - this is a tokenizer round-trip tool. Load it only when you want to run the reconstruction demo, and know that like the rest of this pack it's a single-commit, unmaintained research wrapper.