MOSS-Image-Edit-HQ: Free AI Photo Edits From One Sentence, and Why You Draft Fast and Finish Slow
A community add-on turns the free MOSS-Image poster model into an instruction photo editor under Apache 2.0. Its author published every failed version. Here is a two-speed routine built on what those failures show.
Most AI tools publish their wins. The model card for MOSS-Image-Edit-HQ opens its version history with a loss.
Version 1, the card says, "lost the eval gate," meaning it failed the author's own test against a rival add-on from chinoll, which the card says "preserves identity better": its people still looked like the same people. The author kept training. Versions 2 through 6 went up over the following days, each with a line explaining what changed and why the last one was replaced. Version 6 landed on Hugging Face just after midnight UTC on October 26.
That paper trail is unusual, and it is the best guide to using the thing well. MOSS-Image-Edit-HQ is an unofficial add-on for MOSS-Image, the free poster model OpenMOSS released this month. Give it a photo and one sentence ("make the jacket navy," "remove the sign on the left," "fix the blown-out window light") and it returns the edited picture. It is licensed Apache 2.0, the same as MOSS-Image, which allows commercial use. For client work, also check the terms of the training data the card lists.
The card also says, in plain words, where its fast mode falls short. Read that carefully and a routine falls out of it.
What it actually does
Think of it as a retoucher who takes notes in one sentence. You hand over a photo and say what to change. The add-on treats the photo as the thing to keep and your sentence as the change, rather than painting a new picture from scratch.
The card lists the kinds of edits it was trained on: removing things, faces and expressions, environment and style changes, and text. It calls relighting "one of its strengths." It can take up to four reference images, though as you will see, one is where it is strongest.
The later versions are mostly about keeping what you did not ask to change. Version 4 added masks that tell the training which part of the picture was edited. Version 5 described whole scenes in its training captions. Version 6 describes the finished state of each picture and weights the edited region more heavily. The practical result the author is chasing: the jacket changes colour, and the face, the background and the light on the wall stay put.
Why two speeds
There is a separate free add-on, lodestones' 8-step speed add-on, labelled a work in progress, that cuts the number of passes the model makes from 50 to 8. The card says this stack runs roughly six times faster.
It also says what you lose. The 50-pass quality setting "wins clearly on final edits." At 8 passes, "object additions weaken" (the card's own example is a cat that "barely lands") and "a red cast creeps in."
That is the whole routine in two sentences. The fast mode is a terrible place to judge colour or added objects, and an excellent place to find out whether the model understood your sentence at all. Most failed AI edits fail on wording: the model changed the wrong jacket, read "warmer" as orange, or redrew the whole room when you meant one lamp. You want to find that out in seconds, not in full renders.
So you search for the right sentence fast, then spend one slow render on the sentence that worked.
Put this into practice
You need ComfyUI and a graphics card with about 20 GB of memory. The card says plainly that it will not run on 6 GB. If you want to try before installing, the free demo Space runs both a quality mode and a fast mode in the browser.
1. Install. Update ComfyUI to 0.42.3 or newer and install the ComfyUI-Moss-Image-Edit node pack. Load the MOSS-Image SFT model as you would for poster work. Put armV6-moss-edit-lora.safetensors from the add-on's repo into models/model_patches/. (If you run a recent development build of ComfyUI, there is also a native route using the experimental "Text Encode MOSS-Image Edit" node with edit_format set to ai-toolkit.)
2. Load the workflow. Import workflows/armV6_workflow_single.json. Use the single-photo workflow for nearly everything.
3. Write three versions of your instruction. Keep each to one change. Name the object by what it looks like and where it is: "the navy jacket on the man at left," not "his jacket." Say what should stay: "keep the face, background and lighting." Write one plain version, one more specific, one that adds a keep list.
Here is what that looks like for a product shot of a mug on a café table, where the client wants the mug teal instead of white:
- "Make the mug teal."
- "Make the white ceramic mug in the centre of the table teal."
- "Make the white ceramic mug in the centre of the table teal. Keep the coffee, the table, the window light and the background unchanged."
The first is what most people type. The third is what a retoucher would want written on the job ticket. If the short version tints the coffee, the saucer or the whole frame, that is exactly the kind of misunderstanding you want to catch at 8 passes rather than 50.
4. Draft all three at 8 passes. Add the speed add-on, set guidance (CFG, how strictly the model follows your words) to 1, and render each wording with the same seed, the number that fixes the random starting point. You are only checking one thing: did it change the right object in roughly the right way?
5. Pick the winning sentence. Ignore the colour. The draft will look a little soft and may lean red. That is the fast mode, not your edit. Do not approve a skin tone, a brand colour or an added object from a draft.
6. Render the winner slowly. Remove the speed add-on and render at 50 passes, CFG 3, using the card's other settings (the Euler sampler and a shift of 5.0, both set in the workflow's sampler node). If the edit comes out too timid, raise CFG to 3.5 or 4 and render again.
7. Check what should not have changed. Stack the original and the edit as layers in your image editor and set the top layer to Difference. Bright areas show what moved. You want brightness only where you asked for a change. If the face or background lights up, tighten the keep list and repeat from step 4.
On a typical job, that might be three quick drafts and one or two full renders instead of six full renders spent guessing at wording.
A note on reading community add-ons
The version history is worth a minute even if you never install this. It tells you the author tested against a rival and published the loss, lists known flaws by version (version 3, for example, made colours more saturated than the source), and labels untested claims as untested. Compatibility with the other MOSS-Image models, Base and Kromoss, is marked "unverified" rather than assumed.
That is the behaviour you want from anyone whose files you put in a client pipeline. A card that admits what it has not tested is more trustworthy than one with only a gallery.
The honest limitations
Combining photos is weak. The workflow accepts up to four references, but the card says the training used single-photo pairs and that "putting a person/object from one photo into another is a compositing task it wasn't trained for." Do composites in your editor.
The evidence is the author's. Every comparison grid comes from the author's own test set and demo. There are no third-party scores yet.
It needs a big card. About 20 GB of memory, by the card's own estimate. Laptop and 8 GB users are better served by the earlier, lighter add-on or the demo.
It changes fast. Six versions in under four days is great for quality and hard for consistency. Note which version you used on a job, keep a copy of that file, and do not let an update replace it mid-project.
Fast mode lies about colour. Worth saying twice. The red cast and weak object additions in fast mode are documented by the author. Final approval only happens at 50 passes.
Unofficial. The card says it is not affiliated with or endorsed by OpenMOSS. If OpenMOSS ships its own editing model, compare before you commit.
The bigger point
The interesting thing about this add-on is not that a free model can edit photos. Several can. It is that the author told you exactly where the fast setting goes wrong, which hands you a working method: let the cheap pass answer "did it understand me?" and let the expensive pass answer "is it good?"
That split works far beyond MOSS-Image. Almost every AI image and video tool now offers a fast draft mode next to a slow quality mode, and most people use them as "worse" and "better." They are really two different questions.
Try it on one real job this week, something small like a colour change on a product shot. Write three sentences, draft them in seconds, render the winner once, and run the Difference check. If the background stays black and the jacket turns navy, you have a free retoucher that takes notes in English. If it does not, the drafts will have shown you why before you spent the long render.