Load Anima LLLite Model 🛂🅐🅒🅝

If you've moved to Anima - the anime DiT that finally beat Illustrious on prompt comprehension - you've probably hit the one thing it's missing: ControlNet. This node is how you use what little exists. It loads a ControlNet-LLLite model patch trained for Anima and returns it as a ControlNet you can drive with an Apply Advanced ControlNet node. It's the loader half of Anima's structural-control story.

Set your expectations honestly, because the pack does. Anima's ControlNet is kohya's experimental LLLite port plus a handful of sample weights, and users consistently report the effect is moderate - the pose model especially is deliberately weak. This node works and is the right tool; it just can't give you SDXL-grade pose or depth control, because that model doesn't exist for Anima yet.

How it works

Anima is a 2B DiT, not a UNet, so classic ControlNets don't load on it. kohya's ControlNet-LLLite is a lightweight, LoRA-shaped module injected at the model's attention and MLP points instead of a duplicated encoder - that's what a "model patch" is here. This node reads that patch and wraps it as an Advanced ControlNet so the rest of the pack's machinery (scheduling, masks, per-block weights via the Anima custom-weights node) applies to it. Because LLLite patches the model, remember to route the apply node's model_opt output into your sampler.

The inputs and outputs that matter

• model_patch (required) - the dropdown of Anima LLLite files. The retrained-for-base-1.0 options are essentially two: an inpainting model and a mixed model trained on lineart/scribble/grayscale. Older Preview-3-era lineart, depth, and pose patches still load but run at reduced quality on the 1.0 base.

Optional timestep_kf attaches a schedule when this loader isn't feeding an Advanced apply node; with an apply node present it's redundant.

Output is a CONTROL_NET for your apply node.

How to install it

ComfyUI Manager: search ComfyUI-Advanced-ControlNet, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kosinkadink/ComfyUI-Advanced-ControlNet

then restart. The node ships with the pack; the models don't. Download Anima LLLite patches from kohya's kohya-ss/Anima-LLLite repository and place them in ComfyUI/models/model_patches - that's the folder this loader reads. (The regular Advanced ControlNet loader will also recognize them if you drop them in ComfyUI/models/controlnet instead.)

Common issues & troubleshooting

The dropdown is empty. No patch files in ComfyUI/models/model_patches. Download them from the kohya repo and refresh ComfyUI.

The control barely moves the image. This is the known state of Anima ControlNet, not a mistake on your end. The LLLite weights are experimental and moderate by design, pose weakest of all. Community end-percent tips help a little - the third-party regional LLLite work recommends ending control around 0.45 - but if you need dependable pose/depth conditioning, Illustrious remains the better base for now.

Nothing happens at all. LLLite patches the model, so the apply node's model_opt output has to reach your KSampler. Miss that wire and the control is silently inert. For the 4-channel inpainting model, also route your source mask through the Anima LLLite Extras node into a weights node's cn_extras.