MOSS Image vs Krea 2 vs Qwen 2.1: the license decides

Three open image models share this week's ComfyUI downloads, and for an AI influencer they fail in three different ways. Qwen-Image 2.1 can hold a face from up to 10 reference photos, but its license text allows research or evaluation only. Krea 2 is free for commercial use until your company reaches USD 1,000,000 in yearly revenue. MOSS-Image has Apache-2.0 weights and no reference-image input at all.

Earlier comparisons put Krea 2 against Qwen-Image 2.1 (ours is here) and we reviewed MOSS-Image on its own. Nobody has lined up all three on what a character feed needs, so this article does that from the model cards, the license texts and the ComfyUI repack pages we read today. It also says plainly what is missing: no source we found measures face similarity for any of the three under identical settings, and we did not run them ourselves. If you pick Krea 2, the official template read node by node shows its settings and the prompt enhancer that is on by default.

The verdict first

• Choose Qwen-Image 2.1 to explore and lock a face from reference photos, as long as the output stays inside research or evaluation. It is the only one of the three that takes references in the model itself.
• Choose Krea 2 Turbo for a feed that earns money if your company stays under USD 1 million a year: 8 steps, a clear commercial path and a Raw base meant for LoRA training.
• Choose MOSS-Image if you cannot accept a revenue cap or a research-only license, and you are willing to train a character LoRA. It is the smallest at 5B, and the slowest by its own recommended 50 steps.
• Do not choose by looks yet. The only side-by-side test we found covers Krea and Qwen, on one card, with subjective ratings. Run ten renders of your own reference set before you commit.

Side by side

Keeping the same face: references against LoRA

This is the real split. Qwen-Image 2.1 accepts up to 10 reference images and describes identity preservation for people and products on its model card, with showcase examples and no measurements. Krea 2 offers a style reference workflow with one or two images, which steers mood and look, not who the person is. MOSS-Image takes text, tags or a layout.

So two of the three route identity through training. A character LoRA means a reference set, a rented or local GPU and a test pass; our guide to consistent characters and the one on training a character LoRA cover that. One tester of MOSS-Image reported that a character LoRA turned decent photorealism into results he called amazing, which fits the pattern but is a single hands-on report. A LoRA belongs to the base it was trained on, so choose the base whose license you can live with before you spend the GPU hours.

Musetta studioDemo data

The license decides more than the pixels

Qwen's license defines Non-Commercial as "for research or evaluation purposes only" and says in section 2.b that you shall not use the Materials for any commercial purpose without a separate commercial license. It does not separately address generated images, and that gap is exactly what a user asked the Qwen team to settle in a Hugging Face thread that showed no reply when we read it. We are not lawyers; if the output pays your rent, ask Qwen for the answer in writing.

huggingface.coOpen the page

Krea's terms are the opposite kind: clear and capped. Commercial use is allowed only while total company revenue, affiliates included, stays under USD 1,000,000 over the trailing twelve months; above that you stop and need an Enterprise License. You own what you generate and you are responsible for it. MOSS-Image has the plainest weights license, Apache-2.0, but the project page says the pipeline code and third-party components keep their own licenses, and the Qwen3.5 text encoder and FLUX.2 VAE it runs with are separate files you should check yourself.

What ComfyUI downloads tell you

The Comfy-Org repack pages show how many people pulled each model in the last month. The counts cover every file in a repack, including variants, so read them as interest, not as unique users.

Qwen leads by a wide margin despite its license, which says something about how many people are testing rather than selling. MOSS-Image is days old and already past a million. Download counts do not show quality; they show that a lot of people will hit the same ComfyUI problems you will, which is the next section.

Setup differences that bite

• Qwen-Image 2.1 Turbo gave poor results for some users in ComfyUI until they switched from KSampler to SamplerCustomAdvanced with a ManualSigmas node and the model's own schedule.
• MOSS-Image produced desaturated images with odd curl artifacts for one user until the --use-sage-attention launch flag was removed.
• Krea 2 ships fp8, int8, bf16, mxfp8 and nvfp4 Turbo files in its repack, so the first decision is which one your card runs fast. In a community test on an RTX 3090, int8 was 1.78 times faster than fp8; newer cards may not behave the same.
• All three put the biggest download in the text encoder or the model split. Our file size study explains why, and a rented GPU avoids the disk problem.

When each one wins, and loses

Where Musetta fits

Musetta runs on your own ComfyUI and keeps each character's reference photos, so trying a second base model means installing what the workflow is missing, not rebuilding her. It changes none of the licenses above. See how it compares with working in ComfyUI directly, or fix missing nodes first.