Nodes/ComfyUI-WanAnimatePlus/WanAnimatePlus Uni3C ControlnetLoader
ComfyUI Node

WanAnimatePlus Uni3C ControlnetLoader

Load the Uni3C controlnet that steers your Wan clip

By wuwukaka·Created 4 months ago·Updated about a month ago· 409
WanAnimatePlus Uni3C ControlnetLoader
  • compile_args
  • controlnet
model
base_precisionfp16
quantizationdisabled
load_deviceoffload_device
attention_modesdpa

ControlNets are how you stop a video model from doing whatever it wants. For Wan, the Uni3C controlnet family is one of the ways to inject spatial guidance - depth, pose, or other structure - into the diffusion pass, and this node is the WanAnimatePlus-native way to get one into memory. It's the sibling of Kijai's WanVideoWrapper Uni3C loader, re-exported under the pack's prefix so the whole chain stays self-contained.

The mechanism is standard-issue but worth knowing: the loader reads a controlnet checkpoint from ComfyUI/models/controlnet, checks it's actually a controlnet (it validates the controlnet_patch_embedding.weight key), and loads it as a WANVIDEOCONTROLNET object on whatever device you choose. The actual conditioning - applying it to your latents with a strength - is the job of the sibling WanAnimatePlus Uni3C Embeds node; this one is just the "get it into memory" half.

The settings that matter

  • model - pick from the controlnet dropdown; the tooltip points at ComfyUI/models/controlnet.
  • base_precision - fp16 default, with bf16 and fp32 available. fp16 is the usual Wan controlnet choice; drop to bf16 if you see precision weirdness.
  • quantization - disabled, fp8_e4m3fn, or fp8_e5m2. fp8 shrinks the controlnet meaningfully if VRAM is tight, at a small quality cost.
  • load_device - offload_device is the default and the right call for anyone under ~48GB; the tooltip flatly warns against main_device with larger models unless you have the VRAM.
  • attention_mode - sdpa (default) or sageattn, if you've got SageAttention working (the usual Triton dependency caveats apply).
  • compile_args - optional; feed it the output of WanAnimatePlus TorchCompileSettings to compile the controlnet too.

Output is a single controlnet (WANVIDEOCONTROLNET) that wires straight into WanAnimatePlus Uni3C Embeds.

Install

Same pack drill:

cd ComfyUI/custom_nodes
git clone https://github.com/wuwukaka/ComfyUI-WanAnimatePlus.git

Restart ComfyUI, or install via Manager by searching "ComfyUI-WanAnimatePlus". Drop your Uni3C controlnet files into ComfyUI/models/controlnet, keep the original ComfyUI-WanVideoWrapper installed per the README, and don't mix WanAnimatePlus nodes with original WanVideoWrapper ones in the same workflow.

Honest note: if you're running the official SCAIL-2 flow chain, this node isn't part of it - that path uses the Flow nodes. Reach for the Uni3C pair when you're building a WanAnimatePlus wrapper chain with explicit structural control.

CategoryWanAnimatePlus

Inputs (6)

NameTypeDefaultDescription
modelCOMBOThese models are loaded from the 'ComfyUI/models/controlnet' -folder
base_precisionCOMBOfp163 options: fp32, bf16, fp16
quantizationCOMBOdisabledoptional quantization method
load_deviceCOMBOoffload_deviceInitial device to load the model to, NOT recommended with the larger models unless you have 48GB+ VRAM
attention_modeCOMBOsdpa2 options: sdpa, sageattn
compile_argsoptWANCOMPILEARGS

Outputs (1)

NameTypeDescription
controlnetWANVIDEOCONTROLNET