// the find
kijai/ComfyUI-CogVideoXWrapper
A ComfyUI custom node wrapper for CogVideoX, THUDM's open-source video generation model family. It bridges the diffusers-based CogVideoX pipeline into ComfyUI's node graph, supporting text-to-video, image-to-video, vid2vid, ControlNet, LoRA, and several research extensions like Tora and Go-with-the-Flow. Aimed at power users who want to run CogVideoX locally with fine-grained control over sampling parameters.
Ships ready-to-use example workflows in JSON so you can load and run without figuring out node wiring from scratch. GGUF quantization support means you can run the 5B model on GPUs with less VRAM than the full fp16 weights demand. FasterCache and optional SageAttention integration are real speed levers — not just flags that do nothing. Separating the VAE from the transformer load (with fp32 VAE option) shows awareness of the precision mismatch issues that quietly ruin output quality.
The README is a running changelog, not documentation — there is no stable API surface described anywhere, no parameter reference, and the WIP spreadsheet on Google Docs is a fragile substitute for versioned docs. Breaking workflow compatibility multiple times across updates (Updates 6, 7, and 8 all broke old workflows) suggests no stability policy exists; anyone building on this for production tooling will get burned. The codebase is explicitly acknowledged as bad ('it was bad, still isn't great'), and the single 1600-line nodes.py pattern common in ComfyUI wrappers makes it hard to audit what any given node actually does. Dependency on a specific diffusers version pinned in requirements.txt is a landmine when other ComfyUI extensions pull a different version.