ControlNet Custom Weights [Anima] 🛂🅐🅒🅝

This is the fine-grained weights node for Anima's ControlNet-LLLite models. Where the standard Scaled Soft Weights node gives you a softness curve, this one gives you a slider for each of Anima's 28 transformer blocks - 28 separate knobs, one per block. That's a lot of surface, and it exists because Anima isn't a UNet; it's a 2B DiT (built on NVIDIA's Cosmos-Predict2), and its control is injected block-by-block through kohya's LLLite modules. If you want to sculpt exactly where in the network the control bites, this is the tool.

Worth setting expectations up front: Anima's ControlNet support is experimental. kohya ported ControlNet-LLLite to it in 2026 and published sample weights, and several users report the effect is moderate even at full strength - the pose model especially is deliberately weak. This node is genuinely useful, but it's polishing a control layer that's still one person's reference implementation, not a mature pipeline.

How it works

Anima's LLLite injects a control signal at each of the model's 28 blocks. This node produces a weight for every one of them, so you can zero out blocks that are doing more harm than good, boost the ones that carry the structure you care about, or flatten the whole thing to a uniform value. It's the same "custom weights" concept as the ControlNet and T2I-Adapter versions, sized to Anima's block count. The optional uncond_multiplier additionally controls how much the weights affect the unconditional pass.

The inputs and outputs that matter

The required inputs are block_0 through block_27, each defaulting to 1.0 (range 0 to 10). Leaving them all at 1.0 is a uniform, unmodified application - a fine starting point. You then adjust individual blocks to redistribute the control. In practice you won't hand-tune all 28 from scratch; you'll start flat and pull specific blocks down when the control is overcooking part of the image.

Optional uncond_multiplier (default 1.0) sets the effect on the negative pass. Optional cn_extras is where you pipe in the mask from the Anima LLLite Extras node when you're using the 4-channel inpainting model.

Outputs are CN_WEIGHTS - into weights_override on the Apply Advanced ControlNet node, or cn_weights on a Timestep Keyframe - and TK_SHORTCUT, a ready-made Timestep Keyframe carrying these weights.

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 itself has no dependencies, but it only makes sense alongside an Anima LLLite model - load one with the Load Anima LLLite Model node, and place the model_patch files in ComfyUI/models/model_patches (the regular Advanced loader also finds them in ComfyUI/models/controlnet).

Common issues & troubleshooting

The control is weak no matter what you set. That's Anima LLLite, not this node. The sample weights are experimental and intentionally moderate; the pose model is the weakest. If you need reliable pose or depth conditioning for anime, Illustrious with real ControlNets is still the more dependable choice.

"Incompatible weight type." These weights are for Anima LLLite specifically - don't feed them into a standard ControlNet or a T2I-Adapter, and don't feed those types' weights into an Anima model.

28 sliders is overwhelming. Start with everything at 1.0 and only touch blocks when you see a specific problem. Uniform is a perfectly good default; per-block tuning is for when you're chasing a particular artifact.