PC: Anima attention Couple Model Patch

Regional prompting - giving different areas of the canvas different prompts so your two characters don't swap hair colors - has one running joke behind it: it breaks on every architecture shift and someone has to rebuild it. Latent Couple and Regional Prompter for SD1.5, Attention Couple for Forge, purpose-built nodes when Anima landed in 2026 because, in the words of the person who wrote one, "the built-in regional conditioning does not support the architecture." This node is Prompt Control's answer for Anima specifically, and its category - promptcontrol/experimental - is an honest label, not modesty.

What it does

Attention Couple works by masking conditioning inside the model's attention layers, not by editing the prompt text - that's the whole distinction between it and the prompt-scheduling side of this pack. Anima runs on a different backbone (a Cosmos-Predict2 diffusion model with a Qwen3 text encoder) than the SDXL-family models most Attention Couple implementations were built against, so a generic attention-couple patch doesn't attach the same way. This node is Prompt Control's Anima-specific version of that patch: it modifies how the model computes attention so that region masks attached elsewhere in your graph (via PCAddMaskToCLIP and IMASK() in your prompt, or similar) actually take effect during sampling on this particular architecture.

It's a model patch, not a conditioning node - it doesn't touch your prompt text or your CLIP output at all. It just returns a modified MODEL that's aware of the region masks by the time it reaches your sampler.

Inputs and outputs

One required input, model, one output, MODEL. That's it - all the actual region and mask configuration happens upstream (masks attached to CLIP, referenced from your prompt text), this node's only job is patching the model so it respects them.

Installing it

ComfyUI Manager: search "ComfyUI Prompt Control", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/asagi4/comfyui-prompt-control

No extra downloads for this node specifically - you'll need Anima itself and its Qwen3 text encoder set up separately, same as any Anima workflow.

Where people get burned

Take the experimental category seriously. New base-model architectures get regional-prompting support built for them reactively, after release, and that support tends to have rough edges the older SDXL-era implementations sanded off over years - expect uneven composition, especially if you hold hard region masks all the way through sampling. The general fix that's worked across every generation of this technique is the same here: stop masking hard region boundaries once composition has formed (early in the steps) and let the base conditioning blend the rest, rather than forcing the mask for the entire denoise. If Anima ships an updated native regional-conditioning path later, that's the kind of change that could make this node redundant - check whether it's still the recommended route before building a big workflow around it.