Regional Prompts (Latent Img2Img)

The img2img version of the pack's masked-conditioning regional node - same as "Regional Prompts (Latent)" but starting from a real image, and same as "Regional Prompts (Attention Img2Img)" but using the strict masking path instead of attention patching. If you want to edit part of an existing image and you're okay with (or actually want) cleanly bounded regions, this is the one.

The mechanism is the conservative one: each regional prompt is encoded separately, gets a mask carved from its box, and the sampler applies each masked conditioning in its own region - ComfyUI's standard conditioning-with-mask path, no hooks involved. Because the mask is doing the isolation, regions stay where you put them, which makes this the right tool when a concept absolutely must not leak into the neighboring box. The tradeoff, same as always with masked regional work: the boundaries show, and on img2img that's doubly true because the source image's own content is fighting the mask edges.

Inputs and outputs

• image - source image, picked from your input folder.
• base_prompt - the everywhere prompt (and prepended to each region's prompt).
• prompt_1 through prompt_4 - per-region prompts.
• clip - your CLIP.

Outputs are image (the source passed through for encoding), conditioning (your regional positive), and width/height (source dimensions). Same shape as the Attention Img2Img node - the difference is purely which regional machinery runs underneath.

When to pick this over the Attention version

The README's recommendation stands: attention-based is smoother and the default for most people. But hard isolation is a real need. If you're swapping out a specific object and any hint of the old object's influence in the edit region ruins it - or if the attention hook path is misbehaving on your model - this is the predictable fallback. Also worth knowing: this variant is genuinely useful for testing, because its behavior is more deterministic, so it's easier to debug "why is my region wrong."

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

Watch denoise strength like a hawk - above ~0.5, img2img re-rolls everything and your unboxed areas stop surviving. Expect (and budget for) seams at box edges; overlap boxes a hair and let the normalized mask blend soften them. If a region ignores its prompt, verify the box is fully inside the image and the prompt isn't empty, since both silently no-op.