Can MOSS-Image Edit Images?

The short answer

The clearest verified route is chinoll/Moss-Image-Edit, a community LoRA that works with MOSS-Image Base or SFT. ComfyUI also merged native support for MOSS-Image edit LoRAs on October 22, 2026. These tools can follow a reference image and text instruction, but they remain diffusion-based generation: faces, lettering, textures, and areas outside the intended change can move.

OpenMOSS has said it is exploring reference-related features, but that discussion is not a promise of an official edit checkpoint. Treat community repositories, workflows, and recommended settings as versioned dependencies.

Path 1: the Moss-Image-Edit LoRA

The community Moss-Image-Edit LoRA applies text instructions to one or more reference images. Its published examples cover color, clothing, lighting, season, and visual-style changes. The model card supports both Base and SFT and provides separate single-reference and multi-reference workflows.

Download the model files

hf download chinoll/Moss-Image-Edit \
  adapter_config.json adapter_model.safetensors \
  --local-dir models/Moss-Image-Edit

hf download OpenMOSS-Team/MOSS-Image-SFT \
  --local-dir models/MOSS-Image-SFT

Import a workflow published for this LoRA rather than adding a generic LoRA node to an unrelated text-to-image graph. The reference image must reach the model using the edit workflow's conditioning path.

Path 2: current native ComfyUI edit support

ComfyUI pull request #17296 merged support for MOSS-Image edit LoRAs into the main project. It adds reference-latent handling for the chinoll workflow and edit LoRAs trained with ai-toolkit.

1. Update ComfyUI to a current build containing the merged edit support.
2. Use the official MOSS-Image model loaders and an edit workflow that matches your LoRA.
3. In Text Encode MOSS-Image Edit, select the edit format required by the workflow; the chinoll guide uses Moss-Image-Edit.
4. Start with the published SFT single-reference workflow before adding more images or changing sampling settings.

The Chenkin-x ComfyUI repository remains useful for workflows, installation notes, and a custom-node fallback. Its maintainer explicitly recommends native ComfyUI support for current installations. Base BF16 and multi-image graphs are included, but the repository says their effects still need evaluation.

Path 3: Ostris edit nodes for ai-toolkit LoRAs

ComfyUI-MOSS-Image-Ostris-Edit is an optional MIT-licensed node package for edit LoRAs trained with ai-toolkit. It is not a drop-in replacement for the chinoll workflow.

cd ComfyUI/custom_nodes
git clone https://github.com/ostris/ComfyUI-MOSS-Image-Ostris-Edit.git

The package requires a ComfyUI build with native MOSS-Image support and the Qwen3.5 vision weights used by its text encoder. Its example graph uses 30 steps, CFG 3, skip-layer guidance on layer 22, and a model patch configured for the LoRA's reference-latent method. Treat those values as an example for that workflow, not universal edit settings.

Choose an editing path

There are other community inpainting graphs and shared workflow packs, but their dependencies and claims change quickly. This page does not present an unverified Civitai or forum workflow as an official MOSS-Image feature.

The JSON layout alternative: control before generation

MOSS-Image JSON layout is not image editing. It guides a new generation by describing regions for text and objects with items, bbox, yxyx, and abs_centralized. It can help establish composition before generation, but it does not guarantee an exact coordinate, font, color, or unchanged pixel region.

If your real goal is to position menu text, a poster heading, a product label, or multiple visual elements, build the layout first and generate a new image. That path works in the hosted studio and does not require a local GPU.

Limits and when to use a dedicated editor

Reference-image generation is useful when you want a creative reinterpretation that follows an instruction. It is a poor fit when the task demands deterministic pixels or a specialized computer-vision operation.

• Identity-critical face changes: use a tool designed for identity preservation and review consent and usage rights.
• Background removal: use a segmentation or matting tool rather than regenerating the whole image.
• Exact text replacement: use a layer-based editor when spelling, typography, and unchanged surrounding pixels are mandatory.
• Upscaling: use a dedicated super-resolution model; an edit workflow is not an upscaler.
• Masked object replacement: prefer a documented inpainting model if the untouched region must stay as stable as possible.

Always save the source image, prompt, seed, workflow JSON, LoRA version, and ComfyUI revision. Compare at full resolution and proofread any generated lettering before publishing.