Regional Prompts (Latent + Image Input)

This is the sixth and strictest node in the Steaked-nodes regional family, and if you've been following along you already know what it is: the masked-conditioning (latent) regional method, fed by an IMAGE wire instead of a file picker. You get hard, well-isolated regions on any image that flows into it - no attention patching, no bleeding, and the seams that come with the territory.

The mechanism is ComfyUI's standard conditioning-with-mask path. Each regional prompt is encoded separately, given a mask carved from its box on the source image, and the sampler applies each one inside its own region while the base prompt fills the unmasked remainder. Because the mask does the isolation, a concept in box one stays in box one - that's the whole appeal. It's the deterministic, easy-to-reason-about member of the family, which is a real virtue when you're debugging a regional composition.

Inputs and outputs

• image - the IMAGE wire; anything upstream that outputs pixels works.
• base_prompt - the everywhere prompt, also prepended to each region's prompt.
• prompt_1 through prompt_4 - per-region prompts.
• clip - your CLIP.

Outputs: image (source passed through), conditioning (your regional positive), width, height. Marked as an output node, so it can preview directly.

Why pick this one

Choose it when isolation beats smoothness. If a region's content absolutely must not leak into its neighbor - or when you want the predictable masked behavior to iterate on - this is the node. The wire-fed image input is the bonus: it lets you drive the same regional edit across a changing source (batch, crops, prior outputs) without touching a dropdown. And if you want the same hard isolation but a file-picker source instead of a wire, the Latent Img2Img variant is your file-based twin.

Install

Part of Steaked-nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/StealthNinja1O1/Steaked-nodes

Restart, or ComfyUI Manager → "Steaked-nodes". Nothing extra to install.

Common issues

Three things to keep in mind, all familiar by now. Keep the incoming image resolution consistent, because box geometry is relative to the source and the same boxes land differently on different sizes. Keep denoise strength moderate or the whole canvas re-rolls. And expect seams at box edges - that's the cost of the isolation this node exists to provide; nudge boxes to overlap slightly (they're normalized) to soften the boundary. If a region does nothing, its prompt is empty or its box missed the image.