// the find
kijai/ComfyUI-HunyuanVideoWrapper
ComfyUI wrapper nodes for Tencent's HunyuanVideo text-to-video and image-to-video model. Aimed at people running local video generation who want more control than the official ComfyUI implementation offers — particularly around quantization options, memory management, and early access to features like context windowing and IP2V. Not a beginner project.
Manual memory management (block swapping, explicit VRAM controls) that the native implementation still lacks. FP8 scaled weights support with a bundled scale map, getting close to bf16 quality at a fraction of the memory cost. Good breadth of example workflows covering t2v, i2v, vid2vid, LoRA, and keyframe control — useful for understanding what's actually possible. The IP2V approach (image as prompt conditioning via VLM rather than as first frame) is a genuinely different take that diffusers-style pipelines don't expose cleanly.
The maintainer has publicly stated they're mostly stopping work on it in favor of the native ComfyUI implementation — the README says so twice. That's not a knock on the author, but it means you're adopting something in maintenance mode. Installation is still messy: multiple model files across different HuggingFace repos with partly manual placement and partly auto-download, and the LLaVA text encoder is a 8B model you might not want sitting on disk. The IP2V feature is explicitly labeled experimental and the image_token_selection_expression tuning instructions read like someone figuring it out in public. No tests anywhere in the tree.