Load Wan Fun Model

This is the Wan 2.1-generation sibling of LoadWan2_2FunModel - same idea (PAI's retrained checkpoints that add start/end-frame prediction and control conditioning on top of a base Wan model), one generation older. Wan-Fun on 2.1 is actually where this whole line started: the "Wan-Fun models - start and end frame prediction, controlnet" announcement thread is what put this repo on the community's radar back in early 2025, months before Wan 2.2 or the "Wan2.2-Fun" checkpoints existed. If you're starting fresh in 2026, LoadWan2_2FunModel is the newer, generally-stronger option - reach for this one if you're specifically working with an existing Wan 2.1 pipeline or want the smaller, faster 1.3B tier this family offers that the 2.2-Fun line doesn't.

How it works

Identical loading mechanics to the rest of the pack: load the checkpoint according to its config, apply the GPU memory strategy, output a funmodels bundle for the samplers downstream.

The inputs that matter

• model - ten checkpoints spanning two sub-generations: the original V1.0 line (Wan2.1-Fun-1.3B/14B-InP, -Control) and the improved V1.1 line, which adds camera control (Wan2.1-Fun-V1.1-1.3B/14B-InP, -Control, -Control-Camera). Default is Wan2.1-Fun-1.3B-InP, the lightest option in the list. Per the README, the 1.3B tier runs about 19GB on disk and the 14B tier about 47GB.
• model_type - Inpaint (for -InP checkpoints - this pack's term for start/end-frame prediction, not masked inpainting) or Control (for -Control/-Control-Camera).
• config - only one choice, wan2.1/wan_civitai.yaml. Nothing to get wrong here, unlike the newer loaders.
• GPU_memory_mode - model_full_load, model_cpu_offload (the recommended default per the README), model_cpu_offload_and_qfloat8 (quantizes the transformer to fp8 for more headroom), or sequential_cpu_offload (slowest, lightest on VRAM).
• precision - fp16 or bf16.

Output: funmodels, feeding into WanI2VSampler or WanT2VSampler.

Installing it

Search ComfyUI Manager for VideoX-Fun first. Community reports from this pack's early adopters describe it not always being indexed there, so if search comes up empty:

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

Restart ComfyUI, then download the full HuggingFace repo for your chosen checkpoint into ComfyUI/models/Fun_Models/<exact-name>/. One user working through this exact install described it plainly: clone the whole HF repo, with all its subfolders, into that path - not just the model weights file.

Common issues

The 1.3B tier feels underwhelming next to 14B. That tracks with the broader Wan community's experience of the 1.3B weight class generally - it's the low-VRAM, fast-iteration option, not the quality ceiling. Use it to dial in prompts and settings, then switch to 14B for a final render if you have the VRAM.

Empty model dropdown. The usual folder-naming issue: the checkpoint directory under Fun_Models/ must match the dropdown string exactly and contain the full downloaded repo.

model_type mismatched to the checkpoint. -InP is Inpaint, -Control and -Control-Camera are Control. Get this wrong and the load succeeds but the sampler output is garbage, because the conditioning pipeline doesn't match what the checkpoint was trained on.

Should you use this over LoadWan2_2FunModel? Only if you have a specific reason to: an existing 2.1-Fun workflow, VRAM tight enough that you want the 1.3B tier, or you're chasing camera-control behavior you've already validated on this generation. For new work, Wan 2.2 is the stronger base model across motion, emotion, and prompt adherence - the community consensus that formed around it applies here too, even filtered through the Fun retraining.