🌀 BV Regional Anima LLLite

Anima's regional attention patch gets you character separation, but it's pure prompt-to-region binding - it never sees your geometry the way a ControlNet does. BV Regional Anima LLLite is the pack's answer to that second gap: it renders your regional document to a solid-color control image, loads a local Anima LLLite MODEL_PATCH, and applies the learned layout guidance through ComfyUI's native model-patch runtime. Prompt binding plus actual geometry, in one node.

The LLLite story matters here because of where Anima sits in the ecosystem. The KB's Anima guide is blunt: Anima's ControlNet line is still experimental LLLite sample weights, "not yet load-bearing" - but a community regional-control LLLite (Sen-sou/Anima-LLLite-Regional-Controlnet) got real traction, with reviewers finding it "really adheres to the masks quite accurately for such a small model." This node wires that capability into the BV workflow.

How it works

The chain is: BV document → render regions as a solid RGB control image → load LLLite weights → patch the Anima model during the sampling interval you choose. The control image contains exactly white plus one solid color per enabled region, with overlaps resolved deterministically (the highest-priority region wins). Your editor's display colors and preview opacity don't leak into it - what you see in the editor is authoring cosmetics; what the LLLite sees is a clean, stable layout map.

The node takes model, regional, and model_patch_name (an enum listing the LLLite patches in your ComfyUI/models/model_patches/ folder - the list is empty until you put a file there), plus strength (default 1.0, and notably it can go negative), start_percent (0.0), and end_percent (1.0). Outputs: patched_model, control_image (so you can see exactly what the LLLite consumed), and legend_json (a machine-readable map of color → region).

The model file you have to get yourself

This is the one node in the pack that needs an external download. The pack does not bundle or fetch the weights. Get the adapter from Sen-sou/Anima-LLLite-Regional-Controlnet on HuggingFace and drop it into:

ComfyUI/models/model_patches/

Then restart ComfyUI. If model_patch_name shows nothing to pick, that's why - the folder's empty or the restart hasn't happened. The node also requires a current ComfyUI build with ModelPatchLoader and AnimaLLLiteApply support, so an ImportError at load means "update ComfyUI," not "reinstall the pack."

Settings and the A/B

The verified reference uses Strength 1.0 over the full 0.0–1.0 interval; the README notes shorter intervals and lower strengths intentionally give the model more freedom. The pack's controlled A/B - same seed, model, prompts, and document, only the LLLite patch differing - showed the active patch keeping both subjects on a more comparable scale and depth plane, while the bypassed render kept regional character binding but ran freer on layout. One procedural trap the docs call out: the captured workflow showed the KSampler on randomize; switch it to fixed before you trust any before/after comparison.

Install

Search BV Node Pack in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/BlackVortexAI/bv_nodepack.git

Restart, hard-refresh with Ctrl + F5, then add the model patch as above.

The honest take

If Anima is your main model and "two characters, controlled scale and placement" is a recurring ask, this is the most capable tool in the pack for it - regional text binding and learned layout in one node. Just treat the LLLite as guidance, not segmentation: it won't associate prompt text with a color on its own, guarantee object boundaries, or force an interaction. It tightens the layout; you still write the prompts that make the picture mean something.