Kromoss: A Sharper Fix for MOSS-Image, the Free Poster Model, Three Days After Launch

A community retrain promises to fix the blur in MOSS-Image's starting-point version, keeps the licence that allows client work, and adds an eight-pass drafting mode. Here is what its own numbers prove, and what they leave out.

Open image models now get their best version from somebody other than the people who made them, and it takes about three days.

MOSS-Image went public on October 16. It is an open model built for posters, title cards and anything with a lot of words on it, and it takes a layout as its prompt: you draw boxes, say what goes in each one, and type the exact headline you want printed. OpenMOSS shipped it in two versions. The polished one, SFT, is what most people run. The other, Base, is the starting point people are meant to build on, and in practice its pictures came out soft.

On October 19, at 01:34 in the morning UTC, a trainer who goes by lodestones posted Kromoss. The model card's own description is short and specific: it "mainly fixes the Base model's blurriness." It keeps the Apache 2.0 licence, so you can use what it makes in paid work. It comes with an add-on that renders a draft in 8 passes instead of 50.

That is a strong pitch for anyone who designs for a living. It is also one person's work, posted overnight, with one chart of evidence. Both of those things are true, and the useful question is how to tell, on your own brief, whether it earned the hype.

Why a fix to the "starting-point" version matters

Most creators never touch a Base model. You run the finished version and move on. So why care that someone sharpened the raw one?

Because Base is where everything else gets built. Style add-ons, faster versions and custom training usually start from the least-processed version of a model, since the polished one has already been pushed toward a particular look. If Base produces soft images, every add-on built on it inherits that softness. A sharper Base is a better foundation for the whole ecosystem that forms around a model in its first month.

Kromoss is a full fine-tune, meaning further training of an existing model rather than a new one: lodestones kept training every part of MOSS-Image-Base on 256,000 images at 1024 pixels, and the released file is the result of 4,000 training updates. It uses the same text reader and the same image decoder as MOSS-Image-SFT, with the same plumbing. The card calls it a drop-in replacement. In plain terms, any setup that already runs MOSS-Image should load Kromoss by swapping one file.

What the numbers say, read carefully

The card includes one comparison, on 100 standard test prompts (PartiPrompts), scored with CLIP, an automatic measure of how closely an image matches the words that asked for it. These are the author's own figures:

• MOSS-Image-Base at 50 passes: 27.19
• MOSS-Image-Base with the 8-pass add-on: 27.53
• Kromoss at 50 passes: 26.88
• Kromoss with the 8-pass add-on: 27.82

Read that list the way a client would read a quote. The headline figure, 27.82, is the best score in the table. But the fair comparison is against Base with the same add-on, 27.53, and the gap there is small. Kromoss on its own, at the full 50 passes, actually scores below the model it set out to fix.

Two more details matter. The Kromoss rows were measured at update 2,000 of training, and the file you download is update 4,000, so the published scores describe an earlier version than the one shipped. And CLIP measures whether the image matches the prompt. It says almost nothing about sharpness, which is the thing Kromoss claims to fix.

None of that means Kromoss failed. It means the chart cannot tell you whether it succeeded. The only test that answers the blur question is your eyes, at 100% zoom, on the kind of image you actually make.

The licence is the reason to bother

Plenty of open image releases this autumn came with strings. Some are research-only, so you cannot use the output commercially at all. Others, like FLUX 3 Image Klein, allow free commercial work only for businesses under a revenue ceiling (US$5 million a year, as reported at launch). MOSS-Image stood out because both of its versions are Apache 2.0, a licence that lets you sell, modify and redistribute.

Kromoss keeps that. The card states the weights and the bundled code "are released under the Apache License, Version 2.0." For a freelancer or a small studio, that removes the paperwork question from the decision. If the images are good enough, you can put them in front of a paying client.

There is one gap in that story, and I will come back to it under limitations: the card does not say what the 256,000 training images were.

Put this into practice

Here is a test you can run in an afternoon, assuming you already have ComfyUI set up with an NVIDIA card. MOSS-Image's README lists an NVIDIA card with CUDA, I found no Mac instructions, and no hosted provider is listed on the Kromoss page yet.

1. Update ComfyUI. The card says native MOSS-Image support is in current ComfyUI. If the text-reader loader does not offer a moss_image option, you are on an old version.

2. Download three files from the Kromoss page. The model goes in your diffusion_models folder (kromoss_bf16.safetensors), the text reader in text_encoders (qwen3.5_9b_bf16.safetensors), and the image decoder in vae (flux2-vae_small.safetensors). If you already run MOSS-Image, you have the second and third. Grab 8-steps-dmd-lora.safetensors for your loras folder as well.

3. Pick one real job. Not a cat, not a test prompt. Use a poster or title card you would actually deliver: a headline, a subhead, a date line, an image area. If you have a MOSS-Image layout file from last week, reuse it.

4. Render it three ways with the same seed. First, MOSS-Image-SFT at its defaults. Second, Kromoss at 50 passes, guidance 3, the card's defaults. Third, Kromoss with the 8-pass add-on loaded at strength 1, 8 passes, guidance 1, Euler sampler with the simple schedule.

5. Compare at full size, in three places. Look at the small text first, since soft type is where blur hurts a deliverable most. Then the edges of the main subject. Then the background texture. Note which one you would hand over.

6. Check placement separately. Kromoss is a fix for sharpness. Nobody has shown that it still puts the headline where your layout box says. Lay the three renders over your comp and check positions.

Two practical rules from the card are worth following exactly. Load the 8-pass add-on as an add-on, never baked permanently into the model file: the card warns its changes are so small that the rounding involved in merging can erase them. And treat 8-pass and 50-pass renders as different pictures. A draft at 8 passes tells you whether a layout works. It does not preview what the 50-pass final will look like.

Where it falls short

The evidence is thin. One chart, one measure, one author, scored on an earlier checkpoint than the file you download. The gain over Base with the same add-on is about three-tenths of a point on a measure that does not track sharpness.

Layouts were not tested. The whole reason creatives care about MOSS-Image is its ability to place exact text in exact boxes. The Kromoss card shows a cat and a general test set. It does not show a single layout. Until someone runs your kind of poster through it, assume placement could drift.

The training images are undescribed. The card says 256,000 images at 1024 pixels, with smaller images dropped. It does not say where they came from. The Apache licence covers the model file; it cannot tell you what the model learned from. If your clients ask about provenance, you cannot answer for this one yet.

It is a day old and unhosted. No hosted provider is listed on its Hugging Face page, so trying it means a full download. The card gives no memory figures. And the first version of a community model is rarely the last; check the page for updates before committing a project to it.

It needs NVIDIA hardware. Like MOSS-Image itself, this is a ComfyUI-or-Python job on a graphics card. The README asks for an NVIDIA card with CUDA, and I found no Mac instructions.

What this asks of you

The way open models arrive has changed. Launch day gives you a rough version, and the community spends the next week finishing it. Kromoss is a good example of the pattern: a specific complaint, a specific fix, posted fast, under a licence that lets you work.

The cost of that speed is that you become the quality check. The author did their part by publishing honest settings and a comparison. The part they could not do is test your poster. So run the three-way test on one real job this week. If the small type reads cleaner and the headline still sits in its box, you have a sharper free tool for client work. If not, you have lost an afternoon and kept your current setup, and that is a fair trade for finding out before a deadline does it for you.