// the find
kijai/ComfyUI-WanVideoWrapper
A ComfyUI custom node wrapper for WanVideo and its growing ecosystem of derivative models — SkyReels, VACE, Phantom, ReCamMaster, and a dozen others. Targets people who want to experiment with new video diffusion releases before they land in ComfyUI core, on hardware that can't hold a 14B model entirely in VRAM.
Block-swapping with async prefetch is genuinely well thought out — offloading transformer blocks to RAM while prefetching the next ones during inference is the right approach for consumer GPUs. The LoRA-as-buffer change (attaching weights to modules so they ride the block swap) is a real improvement over the previous load-from-RAM-every-time approach. GGUF model support in the main loader is practical and not common in ComfyUI wrappers. The example_workflows directory covers almost every supported model variant with ready-to-import JSON files, which saves hours of node-wiring.
The README is a changelog masquerading as documentation — actual setup instructions are buried under memory management advisories, and there is no explanation of which nodes do what or how they connect. With 30+ supported models jammed into one repo (SkyReels, Phantom, ATI, HuMo, MultiTalk, LongCat, Ovi, etc.), the codebase is sprawling and there is no indication which features are stable vs. actively broken. The author explicitly says 'this is my personal sandbox' and blocks issue reporters, which is honest but means support is effectively nonexistent. Windows + torch.compile VRAM spikes are a known and partially unresolved issue that will silently kill generation runs.