LongCat Video ComfyUI Guide: Setup, VRAM and Continuation

LongCat-Video is Meituan's 13.6B foundation model for text-to-video, image-to-video, and video continuation. The official repository is the source of truth for the model and Python inference. ComfyUI-WanVideoWrapper is a separate community integration that exposes LongCat nodes, offloading, checkpoint formats, and example workflows.

The fastest reliable path is to make one small result first. Downloading every checkpoint, LoRA, VAE, and workflow before confirming the model fits your use case creates a large failure surface and does not improve the output.

Three paths, three different contracts

Do not move commands or checkpoints blindly between these paths. A file accepted by a community loader is not necessarily an official upstream artifact, and an official Diffusers-style directory is not necessarily a single ComfyUI checkpoint.

Install the community wrapper with a rollback point

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper.git
cd ComfyUI-WanVideoWrapper
pip install -r requirements.txt
git rev-parse HEAD

Save the commit hash beside the workflow JSON. The maintainer explicitly describes the wrapper as work in progress and notes that memory behavior can change when model code, LoRA loading, PyTorch, Triton, or block swapping changes. Upgrade in a second folder or git worktree; keep the last working commit until the same sample passes.

Start from one known workflow

1. Update ComfyUI and the wrapper to a tested pair of commits.
2. Choose a LongCat example from the wrapper repository, not an anonymous download mirror.
3. Put the text encoder, diffusion model, VAE, and optional LoRA only in the folders expected by that workflow.
4. Restart ComfyUI and resolve every red node before queueing.
5. Run the example's own input and dimensions before substituting personal media.

If a node reports a missing model, stop there. Renaming an unrelated file to match the dropdown hides the cause and usually fails later with a shape or architecture error.

First image-to-video test

Use a clean, fully visible subject and a short motion prompt. Keep the first run small enough to finish quickly, then record:

• wrapper commit and model filename;
• input dimensions and output dimensions;
• frame count, steps, seed, attention backend, and block-swap value;
• peak VRAM, peak system RAM, and wall-clock time;
• whether the result preserved identity, geometry, background, and intended motion.

A useful prompt describes the scene first and the movement second. For example: “A ceramic robot stands on a workbench under soft studio lighting. It slowly turns its head toward the camera while the camera remains fixed.” A bounded instruction is easier to judge than a paragraph containing several scene cuts.

VRAM: use a test ladder, not a magic number

There is no official universal “LongCat works on 8GB” contract. Peak use changes with checkpoint precision, attention backend, resolution, frames, model residency, LoRAs, wrapper commit, and offload settings. Block swap can trade VRAM for system RAM and time; it cannot make resources infinite.

Test in this order:

1. Lower frame count.
2. Lower width and height while preserving the aspect ratio.
3. Use the wrapper's documented offload or block-swap path.
4. Try a supported lower-precision checkpoint from a traceable publisher.
5. Change only one setting, then rerun the same input.

If the machine starts paging heavily, a run that technically avoids CUDA OOM may still be operationally unusable. Record total time and memory, not only whether a file appeared.

Video continuation without the “unlimited” myth

The official LongCat model is pretrained for video continuation and describes minutes-long generation. A practical continuation workflow still generates and joins bounded segments. Every new segment can amplify drift.

Use this acceptance loop:

1. Generate the first short segment.
2. Review its final frames for a stable subject and background.
3. Continue once with a simple action.
4. Inspect the boundary frame by frame.
5. Stop when identity, color, geometry, or motion no longer passes the same quality bar.

Avoid claiming “infinite” or “zero drift.” Those phrases turn a model capability into a guarantee that neither the official repository nor a community workflow can make for every input.

Failure map

Output quality gate

Watch the final video at normal speed and frame by frame. Verify prompt intent, subject identity, anatomy, object count, text, color, camera movement, boundary frames, and playback compatibility. Keep the original seed and workflow JSON with the accepted output. If the job is commercial, confirm model, input, music, voice, portrait, and dependency rights separately.

Primary sources

• LongCat-Video official repository
• LongCat-Video official model card
• ComfyUI-WanVideoWrapper repository and installation notes

If the local setup cost is still unknown, run one hosted video test and compare its accepted result and credit cost with your measured local runtime. That gives you a production decision, not just a successful installation screenshot.