Regional prompting + regional LoRAs for MOSS-Image using the layout bboxes (lightweight multi-pass workflow + small custom node).
retained text
This ComfyUI workflow lets you put up to 5 different character LoRAs into one MOSS-Image scene without them bleeding into each other. It draws the whole scene first with your style LoRA (if you have one), then redraws only the area inside each bounding box using that box’s character LoRA. A small custom node reads the boxes straight from the `CreateBoundingBoxes` node, so when you move a box or change the resolution, the masks follow automatically with no typing coordinates. Just drop the `.py` file into `custom_nodes`, load the workflow, pick your LoRAs, and draw your boxes. Right now it supports **5 different character/loras**, if you need more or less, just bypass the unused ones or draw more from the bbox nodes. Download the workflow and the custom node from [civit.ai](https://civitai.com/models/3024330/regional-character-loras-for-moss-image-using-the-layout-bounding-boxes-two-pass-workflow-small-custom-node?modelVersionId=3434494) Extract the .zip and drop the **moss\_bboxes\_to\_masks.py** into your custom nodes folder and restart comfyUI before dragging the .json workflow into the canvas. **Note 1:** it uses the standard MOSS bbox node which only works in Nodes 2.0 unfortunately but i'm trying to use as little to no custom nodes as possible. **Note 2:** you add any number of your own loras anywhere, it's compatible with SFT and Base as well as all the turbo 4step/8step LoRas, just lower the strengths a little bit if you're stacking a lot of loras. **Note 3:** I made this workflow/node because the standard custom bbox doesn't let me hook loras into each individual bbox while the **Moss freefuse** workflow is really super inaccurate and still tends to bleed a lot. I've been resorting to using Moss Inpainting but this system has so far been a lot more reliable On another note, I really didn't like MOSS at first but it's really starting to become undoubtedly one of my favorite models. works just fine on 10gb vram, could even go smaller if you use gguf.
document fetched 2026-10-29 00:43 · canonical https://reddit.com/r/comfyui/comments/1xhftai/regional_prompting_regional_loras_for_mossimage