Load Wan Model

Every workflow built on VideoX-Fun starts here. LoadWanModel reads a Wan checkpoint off disk, applies your VRAM strategy and precision, and hands the result downstream as a bundle every sampler in this pack expects. If you're wondering why the output socket is typed FunModels even though this specific node loads plain Wan checkpoints, not "Fun"-branded ones - that's just this pack's internal name for its model bundle, used across the whole node set. Don't read anything into it.

What it actually loads

The model dropdown lists four options: Wan2.1-T2V-1.3B, Wan2.1-T2V-14B, Wan2.1-I2V-14B-480P, Wan2.1-I2V-14B-720P. These are the base Wan 2.1 checkpoints from Alibaba's Tongyi lab, not the Fun-branded derivatives (Wan2.1-Fun-*) that VideoX-Fun's own README also ships. That distinction matters when you're picking a sampler to pair this with - WanFunT2VSampler and friends are built around the Fun family, so make sure the checkpoint you load here actually matches what your downstream sampler is expecting.

1.3B is the lightweight tier: usable on 6-8GB cards, noticeably below 14B quality, but fine for testing a workflow before committing to a full render. 14B is the model people actually mean when they say "Wan is crazy" - better motion, better prompt adherence, and the one with the deep LoRA ecosystem behind it. The I2V variants are image-to-video: feed a starting frame and Wan animates forward from it, which the community treats as the default mode because anchoring to a known first frame solves half the consistency problems that plague pure text-to-video.

GPU_memory_mode: the knob that actually matters

This is the field worth reading carefully, straight from VideoX-Fun's own README:

• model_full_load - everything stays on the GPU. Fastest, and the one that needs the most VRAM.
• model_cpu_offload (the default) - the whole model moves to CPU when not actively running. The recommended setting if you've got the VRAM to spare.
• model_cpu_offload_and_qfloat8 - same offloading, plus the transformer gets quantized to float8. Saves more memory; the README is upfront that this "may slightly reduce model performance."
• sequential_cpu_offload - each layer moves to CPU individually rather than the whole model at once. Slowest by a wide margin, but the deepest VRAM savings - this is the one that gets a 14B model running on a card that has no business running it.

14B-class Wan models generally want 12-16GB minimum and are comfortable at 24GB; if you're under that, start with model_cpu_offload_and_qfloat8 rather than fighting model_full_load into an out-of-memory error.

The rest of the fields

config only offers wan2.1/wan_civitai.yaml - a bundled YAML defining architecture hyperparameters for the checkpoint family. Leave it alone; there's nothing to tune here unless you're deep in the repo's training code. precision is fp16 or bf16 - both work fine for inference, bf16 has a slight numerical-stability edge on newer GPUs, fp16 is the more universally supported choice on older ones. Pick whichever your card handles without complaint.

The single output, funmodels, feeds directly into a sampler (WanFunT2VSampler, WanFunInpaintSampler, etc.), or through LoadWanLora first if you want to layer a LoRA on top.

Installing VideoX-Fun

ComfyUI Manager: search "VideoX-Fun" (the older name "CogVideoX-Fun" also turns it up) and install. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/aigc-apps/VideoX-Fun, then restart.

Model files do not go in your usual checkpoints/ or diffusion_models/ folder - VideoX-Fun uses its own directory, ComfyUI/models/Fun_Models/<ModelName>/, and expects a full diffusers-format model folder (not a single .safetensors), matching whatever you download from the README's model zoo table. Budget real disk space: 19GB for the 1.3B tier, 47GB for 14B. The README's own estimate is "about 60GB available on disk" just to get one or two models set up.

Common trip points

Wrong folder structure is the number one issue - if VideoX-Fun's own model doesn't show up or fails to load, check it landed under Fun_Models/ as a full folder, not a loose file. Running 14B on a card under ~16GB without setting GPU_memory_mode correctly is the second - you'll hit an out-of-memory error mid-run rather than at load time, which is confusing the first time it happens. And because this pack ships its own copy of the Wan pipeline rather than plugging into ComfyUI's native model loading, don't expect the checkpoints you've already downloaded for Kijai's WanVideoWrapper or ComfyUI-core's native Wan support to be interchangeable - VideoX-Fun wants its own folder layout.