How to edit MOSS-Image results

1. Load a layout workflow, “MOSS-Image: Layout-Guided Text to Image” or its Base version, from the Template Library or our workflow page. It needs ComfyUI v0.42.0 or newer.
2. Run it once, then set the seed's control after generate to fixed, so every run starts from the same noise.
3. To work on that image, save it, load it with a Load Image node and connect it to the background input of Create Bounding Boxes: the canvas shows it behind your boxes, so you can line them up with what is already there.
4. Change one thing at a time, a word, a description or a box, then run again with the same seed and compare.

Each run generates the whole image again. The fixed seed gives every run the same starting noise, which helps the rest of the picture stay similar, but nothing is locked: other parts can shift too.

In Create Bounding Boxes you draw in pixels from the top-left corner. MOSS-Image Layout Prompt turns each box into the format MOSS-Image was trained on, [y_min, x_min, y_max, x_max] with the origin at the center of the canvas. You only need that format when you write layout JSON by hand for Gradio or Python.

Example on a 1024×1024 canvas: a box whose top-left corner is at x 100, y 50, 300 px wide and 200 px tall. Subtract half the height from each y and half the width from each x:

"bbox": [-462, -412, -262, -112]

ai-toolkit can now train MOSS-Image edit LoRAs, which read up to three reference images along with the prompt, and Ostris's ComfyUI-MOSS-Image-Ostris-Edit nodes run them. You need such a LoRA: train one with ai-toolkit's MOSS-Image Base or SFT [Edit Training] model, or use one someone has published.

1. Clone ostris/ComfyUI-MOSS-Image-Ostris-Edit into ComfyUI/custom_nodes/ and restart. It relies on ComfyUI's native MOSS-Image support (v0.42.0 or newer).
2. Load the edit LoRA after Load Diffusion Model, then add MOSS-Image Ostris Edit Model Patch: without the patch the stock model ignores reference images.
3. Encode the prompt with Text Encode MOSS-Image Ostris Edit and connect up to three images and the VAE, scaling each reference to about 1 megapixel first. The node's example uses 30 steps and CFG 3 with skip-layer guidance on layer 22.

The Comfy-Org text encoder files include the Qwen3.5 vision weights these nodes need. A 1024×1024 reference is 4,096 extra tokens; the kv_cache option, on by default, computes them once instead of on every step, about twice as fast. Turn it off only for a LoRA trained with kv_cache: false.

darask0's MOSS-Image Color-Depth Control is an unofficial adapter that makes generation follow two maps: a color region map, with each object a flat color (the colors are only IDs), and a depth map, where near is white. Either map can be left out. It is still text-to-image from noise: the maps set placement, scale and pose, and your prompts decide what fills them.

• ComfyUI support, the MOSS-Image Color/Depth Control node, was merged on Oct 19 after v0.42.1, so it needs a nightly build; on the portable version, run update_comfyui.bat.
• Download moss_image_color_depth_control.safetensors from Comfy-Org/MOSS-Image (repackaged on Oct 20) into ComfyUI/models/loras/, load it with a LoRA loader, then connect the maps to the MOSS-Image Color/Depth Control node.
• For maps made from real photos keep the control on to the end (1.0); for 3D mannequins and other renders, end it at 0.5 so the model finishes the surfaces.
• Give each color its own region prompt: with only a scene prompt, a mannequin-shaped map comes out as a mannequin.

To rework a picture you already have, make a depth map of it (the adapter was trained on Depth Anything V2 maps) and paint flat-color regions over it, then describe the new content.

The adapter is Apache-2.0. Its training depth maps were made with Depth Anything V2 Large, which is CC-BY-NC-4.0.

• As of Oct 22, 2026, OpenMOSS has no inpainting, masked-edit or instruction-edit model; editing from reference images exists only as community edit LoRAs.
• OpenMOSS says simple fine-tuning can add reference-based generation and GIF generation to Base: “Stay tuned.”