Omost Layout Cond (OmostDenseDiffusion)

Omost Layout Cond (OmostDenseDiffusion) is the regional backend the original Omost repo actually uses, and it's the one to reach for when the ComfyUI-Area node's masked attention keeps letting regions bleed into each other. It's lllyasviel's Method 3 - attention score manipulation. Instead of merely masking attention, it actively rescales the attention scores so each region is encouraged inside its mask and discouraged outside it. Tighter bounds, less leakage between subjects, at the cost of one more install.

That install requirement is upfront: this node does not work on a stock install. It imports huchenlei/ComfyUI_densediffusion, the companion pack by the same author, and errors loudly if it's missing. Install both this pack and ComfyUI_densediffusion - ComfyUI Manager will usually find the dependency once you search for it, or clone it into custom_nodes - restart, and you're set.

Three inputs, all required: model (MODEL, the base model to patch), canvas_conds (the layout), and clip. Two outputs: MODEL - the patched model, which you send to the sampler instead of your original - and CONDITIONING. The mechanism, per the source: each canvas region gets its prompt encoded with the greedy-bag encoder, a mask is attached at strength 1.0, and the conditions accumulate into the patched model via Dense Diffusion's add-cond, then it applies. You get one patched model carrying every region's constraint.

The big caveat, stated in the pack's own README: ComfyUI_densediffusion does not compose with IPAdapter. If your workflow leans on IP-Adapter for identity or style, you cannot have both - pick the ComfyUI-Area backend instead and use its MASK output if you need per-region control. People hit this after wiring a whole graph, so know it before you start.

Everything else is standard Omost plumbing: canvas_conds comes from the chat node or either loader, clip is your model's CLIP. The honest read on which backend to pick: ComfyUI-Area is free and usually fine; this one is noticeably better at holding a subject inside its box, which matters when two characters share a frame and you're fighting attribute bleed - the exact case the regional-prompting literature calls the primary one. If region separation is your whole problem, the extra pack is worth it. If you're just experimenting, start built-in.