Load Wan 2.2 Fun Model

Stock Wan 2.2 does text-to-video and image-to-video well, but it doesn't natively do start/end-frame prediction or take a Canny/Depth/Pose control video - those are things Alibaba's PAI team retrained into a separate line of checkpoints they call "Wan-Fun." LoadWan2_2FunModel is the loader for that line: the Wan2.2-Fun-* family, in both the 14B-active A14B tier and a lighter 5B tier, each shipping in -InP (frame prediction), -Control (conditioning video), and -Control-Camera (adds camera-motion control) flavors.

This is the loader you want if the plain LoadWan2_2Model doesn't have the capability you need - reach for it specifically when you're doing controlled generation, not open-ended T2V/I2V.

How it works

Same load-and-configure pattern as the rest of this pack: pick a checkpoint, VideoX-Fun builds the transformer stack from the matching YAML config, applies the memory strategy, and returns a funmodels bundle for the next node. What changes here versus the plain Wan 2.2 loader is which checkpoints are available and what the model was actually trained to do with its conditioning inputs.

The inputs that matter

• model - six checkpoints: Wan2.2-Fun-A14B-InP, -Control, -Control-Camera, and the same three for Wan2.2-Fun-5B. Default is A14B-InP. Per the README's model zoo, every A14B-tier checkpoint here is 64GB, and every 5B-tier one is 23GB - budget disk accordingly.
• model_type - Inpaint or Control. Pick Inpaint for the -InP checkpoints (this pack's term for start/end-frame image-to-video, not masked inpainting) and Control for -Control and -Control-Camera checkpoints.
• config - only two choices here, wan2.2/wan_civitai_i2v.yaml or wan2.2/wan_civitai_5b.yaml, matching the A14B or 5B tier respectively - there's no separate t2v config in this family because the Fun checkpoints are trained around image conditioning.
• GPU_memory_mode - model_full_load, model_cpu_offload (the recommended default), model_cpu_offload_and_qfloat8 (adds fp8 quantization), or sequential_cpu_offload for the tightest VRAM.
• precision - fp16 or bf16.

Output: funmodels, into a Wan2_2FunInpaintSampler or Wan2_2FunV2VSampler.

Installing it

Search ComfyUI Manager for VideoX-Fun; if it doesn't turn up - this pack has a history of not always being reliably indexed there - clone it directly:

cd ComfyUI/custom_nodes
git clone https://github.com/aigc-apps/VideoX-Fun
pip install -r VideoX-Fun/requirements.txt

Restart ComfyUI, then pull the full HuggingFace repo for your chosen checkpoint (the whole folder - config files and all, not one .safetensors) into ComfyUI/models/Fun_Models/<exact-name>/. At 64GB for the A14B tier, this is one of the heavier downloads in the whole pack - make sure you actually have the space before you start.

Common issues

model_type doesn't match the checkpoint. This is the number one gotcha across every loader in this pack: -InP needs Inpaint, -Control/-Control-Camera need Control. Get it backwards and the node will load without erroring, then produce nonsense output, because the sampler downstream is now feeding it inputs the checkpoint wasn't trained on.

config set to the wrong tier. There's no cross-checking here - pick _i2v.yaml for an A14B checkpoint and _5b.yaml for the 5B one, or loading will fail outright.

64GB checkpoint, and you don't have the VRAM. These are the same weight class as any 14B-active Wan model. If you're not on a 24GB+ card, go straight to sequential_cpu_offload, or drop to the Wan2.2-Fun-5B tier - it's a fifth the storage size and a fraction of the compute, at a real but honest quality cost.

You're doing Control generation and don't have a control video ready. For Canny/Depth/MLSD you'll need ComfyUI's separate controlnet-aux preprocessors; this pack only ships its own pose extractor (VideoToOpenpose), so pose is covered natively but the other control types are not.