One reference image can show an AI editor what a subject looks like. A second can show the outfit or visual direction you want. The difficult part is telling the model which details belong to which image—and which details must not change.
That makes this week's MOSS-Image Edit release worth watching. The community-built adapter brings instruction-based editing to MOSS-Image, with support for up to four reference images. This guide explains what is documented, how to organize a useful reference set, and how to judge the result before using it in a creative project.
What is new in MOSS-Image Edit?
The developer's model card describes MOSS-Image Edit as a LoRA: an adapter used alongside a base model, rather than a standalone image generator. It supports MOSS-Image-Base and MOSS-Image-SFT and lists color, outfit, lighting, season, and visual-style edits as intended uses. The card links single-reference and multi-reference ComfyUI workflows and explicitly says to explain each image's role when supplying multiple references.
ComfyUI Wiki's October 22 coverage reports the community release and native ComfyUI support. The corresponding ComfyUI pull request records the editing-support change. These sources establish a new editing option, not a guarantee of perfect subject preservation or a win over other image models.
Scope: This is a source-based workflow guide, not a hands-on benchmark. MOSS-Image Edit is a community project, not an AI Photo Generator feature announcement. Model availability in any hosted service should be checked separately.
Start with a reference map, not a longer prompt
Before generating anything, assign each reference one job. A clear two-image brief is usually a better starting point than four images with conflicting lighting, poses, and backgrounds.
- Image 1: the anchor. The subject, composition, and details you want to preserve.
- Image 2: the change reference. The outfit, color, or other specific attribute you want to introduce.
- Additional references: only if needed. Give each an explicit, non-conflicting role. More inputs are not proof of better control.
For example, a portrait plus a jacket photograph can form an outfit-editing brief. Two portraits with different people, different poses, and different lighting create a much less precise instruction unless you explain exactly what each contributes. Do not assume that multi-reference support means reliable face swapping, virtual try-on, or accurate garment fit.
Use images you own or are authorized to edit. Get appropriate permission for identifiable people, especially when changing their appearance or placing them in a new context.
A practical two-reference editing workflow
1. Write the acceptance criteria first
Separate the requested change from the protected details. For a jacket concept, the change might be “replace the jacket with the blue jacket from image 2.” Protected details might include the person's face, pose, hands, framing, and original background.
Make the decision testable: the jacket should be blue, its closure should resemble the reference, and the person's face must not visibly change. If a particular logo or garment construction must be exact, plan for manual correction or use real photography rather than assuming the model will reproduce it.
2. Establish a single-reference baseline
Begin with the anchor image and one simple instruction, such as changing a jacket's color. This separates basic editing problems from reference-assignment problems. Keep a copy of the original and record the model, workflow version, prompt, and settings used.
Using image 1 as the source, change only the jacket color from gray to dark blue. Preserve the person's face, hair, pose, hands, jacket shape, camera framing, and background. Keep the original lighting direction.
This is a suggested test prompt, not a promise that every protected detail will remain untouched. Compare the output with the original before adding complexity.
3. Add the second reference and name its job
Once the simple edit is acceptable, add the outfit reference. Avoid a vague instruction such as “combine these.” Tell the editor which image supplies the subject and which supplies the garment.
Use the person and scene in image 1 as the base. Replace the jacket with the jacket shown in image 2. Use image 2 only as the jacket reference; do not copy its person, pose, or background. Preserve the face, hairstyle, hands, pose, framing, and background from image 1. Adapt the jacket lighting to the scene in image 1.
If the result changes the face or imports the second background, simplify the brief and inspect the reference set. A clear garment reference may be more useful than another elaborate sentence. Change one variable at a time so you can tell what helped.
4. Compare candidates against the original
Review at full size and at the intended delivery size. Put the original and edited image side by side; if you have a layer-based editor, an opacity overlay can reveal shifted facial features or a changed silhouette.
- Requested edit: Did the right attribute change?
- Subject preservation: Did the face, proportions, or pose drift?
- Reference separation: Did an unwanted person, background, or accessory carry over?
- Physical plausibility: Do seams, sleeves, hands, shadows, and reflections make sense?
- Delivery accuracy: Are labels, text, and important product details faithful enough for the intended use?
Reject a beautiful image that fails the brief. For commercial product imagery, a convincing invented detail can be more damaging than an obvious visual glitch.
Two more briefs worth testing
Seasonal scene variation
Use one landscape or exterior image to explore seasonal art direction without also requesting a new camera angle.
Transform the vegetation and seasonal atmosphere in image 1 from summer to autumn. Preserve the building geometry, path layout, camera position, and framing. Use natural autumn foliage colors. Do not add people, signs, or new structures.
Check windows, rooflines, and paths carefully. Treat the result as an edited illustration or concept, not evidence that a real location looked that way.
Lighting direction study
Test a lighting edit separately from an outfit or style change. That makes failures easier to diagnose.
Change the lighting in image 1 to soft, warm late-afternoon light from camera left. Preserve the subject, pose, clothing, composition, and background objects. Keep shadows consistent with the new light direction.
Inspect both the subject and the environment. A warm face with unchanged opposing shadows is not a finished lighting edit.
What to check before trying it in ComfyUI
The model card links the developer's installation guide and example workflows. Start there rather than mixing instructions from unrelated MOSS implementations. Native ComfyUI integration and the developer's separate node pack are different setup paths; model packaging and node expectations can differ.
The October 22 ComfyUI Wiki report notes that native editing support landed after the v0.42.1 tag and describes the editing node as experimental. Check your installed version and the current project instructions instead of assuming that any recent installation includes it. Back up a working production setup and test changes in a separate environment where practical.
A small adapter download does not mean the complete workflow has small hardware requirements: it still needs its underlying model and supporting components. Check current documentation for your chosen setup before committing to a local deployment. This guide does not claim a particular GPU, memory requirement, or generation speed.
Licensing and production limits
The model card identifies the project code as GPLv3. Do not treat that single statement as blanket clearance for every weight file, base model, reference image, or generated output. Review the applicable terms for each component and your intended use, particularly before commercial deployment or redistribution.
For client work, keep the authorized reference files, approved output, prompt, and workflow details together. Label concept imagery appropriately and disclose material AI alterations where relevant. If exact identity, product construction, or factual appearance is essential, manual editing or photography may be the more dependable choice.
Frequently asked questions
Is MOSS-Image Edit a standalone image generator?
No. The published release is an editing LoRA used with MOSS-Image-Base or MOSS-Image-SFT. It requires a compatible workflow and the supporting model files.
How many reference images does it support?
The developer documents up to four. Start with the smallest set that communicates the edit, and describe each reference's role explicitly.
Does using references guarantee an unchanged face or product?
No. References guide the edit; they do not remove the need for visual inspection. Treat protected details as acceptance criteria, not capabilities guaranteed by the prompt.
The takeaway: specify what each image contributes
MOSS-Image Edit's timely contribution is a documented route to instruction-driven, multi-reference editing. The practical lesson is broader: choose an anchor, assign each additional reference a narrow role, request one clear change, and inspect what should have stayed the same.
For your next creative brief, explore the text-to-image and image-to-image options in AI Photo Generator, using only the controls and models currently available there. Whether you work in a hosted tool or a local workflow, a precise reference map is more useful than a prompt full of competing requests.