Skip to content

Top listTop listImage modelsLicensesComfyUIFLUX 3 KleinMiniMax H3 Image

Best open source image models in 2026 for a recurring face

· 10 min read · 18 sources

Download counts as of , from each site

In short

For a character you post and earn from, FLUX 3 Image Klein is the best open model today: third on the Artificial Analysis open-weights board at 1041, outputs usable commercially under $5M yearly revenue, up to 10 reference photos, and a 10.48 GB smallest kit. MOSS Image is the only Apache-2.0 pick near the top. MiniMax H3 Image scores highest (1066) but its license excludes the US, EU, UK and South Korea. Qwen-Image 2.1 and Fibo 2.0 score well and forbid commercial use.

The best open image model on the leaderboard is not the best one for a character you get paid for. MiniMax H3 Image sits first on the Artificial Analysis open-weights board at 1066, and its license does not grant you anything if you live in the US, the EU, the UK or South Korea. Two of the next three models on that board are licensed for research only.

So we ranked seven open models the way a creator with one recurring face has to: can you sell what it makes, does it take reference photos of your character, what does the smallest ComfyUI kit weigh, and only then how it scores. Every license below was read at its own file on 17 November 2026, and every size comes from the Hugging Face file listing, not a blog.

The ranking: seven open image models for a paid character feed

Musetta is not a model and is not on this list. The order weighs license first, references second, score and VRAM third, because a beautiful image you are not allowed to post is worth nothing to a feed.

  1. 1

    FLUX 3 Image Klein (Black Forest Labs)

    The best balance today: third on the open board at 1041, outputs usable commercially under $5M revenue, up to 10 reference images, and the second-lightest kit.

    Best for: A monetized character feed on a 12 to 16 GB card

    • Elo 1041, inside the 2 to 4 range with Qwen 2.1 and Fibo 2.0
    • FLUX Kommunity License: outputs for commercial use if you and affiliates earn under US$5,000,000 a year
    • Up to 10 reference images; distilled model runs 16 steps
    • 10.48 GB smallest kit (17.20 GB with the prompt upsampler BFL recommends)
    • Creators report hand, eye and body-type regressions against FLUX.2
    • Only outputs are licensed commercially; fine-tunes and redistribution stay non-commercial
  2. 2

    MOSS Image (OpenMOSS)

    The only Apache-2.0 model in the top group: no revenue line, no territory list, 13 GB kit. It pays for that with 50 steps and no reference input.

    Best for: Creators who want no license ceiling at all, with a character LoRA

    • Apache-2.0 weights, 5.08B parameters
    • 13.00 GB smallest kit in ComfyUI
    • Base checkpoint meant for fine-tuning, so LoRAs carry the face
    • Not on the Artificial Analysis board, so no comparable score
    • No reference images: identity comes from a LoRA
    • 50 steps at CFG 3 by default; backgrounds tend to blur
  3. 3

    Krea 2 Turbo (Krea)

    The biggest LoRA and checkpoint scene of the 2026 models and 8 steps, but commercial use stops at $1M revenue and the license makes content filtering your job.

    Best for: Small creators who live on Civitai realism LoRAs

    • 12.82B parameters, 8 steps
    • 12.78 GB smallest kit (NVFP4)
    • 9 of this week's 20 most downloaded Civitai entries are Krea 2 or Krea 2 Turbo
    • Krea 2 Community License: commercial use only under US$1,000,000 company revenue
    • Section 4.2 requires you to run content filter measures
    • Text to image only; a creator reports flat, blank expressions
  4. 4

    MiniMax H3 Image (MiniMax)

    The highest score on the open board and up to 16 references, behind a license that excludes the four biggest creator markets.

    Best for: Creators outside the US, EU, UK and South Korea and room for a 29 GB kit

    • Elo 1066, first on the open board
    • 6 steps, up to 16 reference images, around 16 MP output
    • Comfy-Org repack loads in core ComfyUI 0.45.0
    • No license grant in the EU, UK, Republic of Korea or USA
    • Authorization needed above US$1M annualized revenue; must display MiniMax-H3-Image on a commercial UI
    • 29.41 GB smallest kit, the heaviest here
  5. 5

    Z-Image Turbo (Tongyi-MAI)

    A year old and 104 points behind the leader, but Apache-2.0, 8 steps and the lightest kit on the list. The fallback that always works.

    Best for: 8 GB cards and anyone who wants zero license risk

    • Apache-2.0, 6.15B parameters
    • 8.33 GB smallest kit
    • 8 function evaluations, guidance 0
    • Elo 937, a generation behind
    • No reference input; consistency needs a LoRA
  6. 6

    Qwen-Image 2.1 (Alibaba)

    Second on the board at 1047 with 10 references and a 14.25 GB kit. Licensed for research or evaluation only, so it cannot run a paid feed.

    Best for: Testing looks and prompts before you pick a commercial model

    • Elo 1047
    • Up to 10 reference images; generation and editing in one model
    • 7.12B parameters, INT8 at 7.26 GB
    • Qwen Research License: Non-Commercial means research or evaluation only
    • Edits reusing the reference's seed come out etched with noise
  7. 7

    Fibo 2.0 dev (Bria)

    Tied with Klein and Qwen on score, but CC BY-NC, gated, 50 steps and no Comfy-Org repack. Last for a creator, useful for brands who buy Bria's license.

    Best for: Teams that will sign a commercial agreement with Bria

    • Elo 1037
    • Up to 4 references; Bria ComfyUI nodes exist
    • CC BY-NC 4.0: non-commercial only
    • Gated repo, 8.13B parameters, about 16.25 GB in BF16 for the generator alone

The numbers side by side

One table, three sources: the score from Artificial Analysis, the license from each model's own file, the kit from the Comfy-Org file listings (smallest generator plus smallest text encoder plus VAE, our sums in decimal GB).

ModelOpen-board EloCan a creator sell outputs?ReferencesStepsSmallest kit
MiniMax H3 Image1066Not in US, EU, UK, KR; auth above $1MUp to 16629.41 GB
Qwen-Image 2.11047No, research or evaluation onlyUp to 1040 (Turbo 8)14.25 GB
FLUX 3 Image Klein1041Yes, under $5M revenueUp to 1016 (base 50)10.48 GB
Fibo 2.0 dev1037No, CC BY-NC 4.0Up to 450no repack
Z-Image Turbo937Yes, Apache-2.0None88.33 GB
Krea 2 Turbon/aYes, under $1M revenue, with filtersNone812.78 GB
MOSS Imagen/aYes, Apache-2.0None5013.00 GB
Read 17 Nov 2026. Kit = smallest diffusion file + smallest text encoder + VAE from the ComfyUI repack. n/a = not on the board.
Open-weights text-to-image Elo, Artificial Analysis
0298.5597.0895.41.2k1.1k1.0k1.0k1.0k1.0k937MiniMax H3 Im…Qwen-Image 2.1FLUX 3 KleinFibo 2.0FLUX.2 devZ-Image TurboUnit: Elo

Source: Artificial Analysis, AA-Image-T2I v2.1 open weights

Read the chart with its error bars in mind: the board puts Qwen 2.1, Klein and Fibo 2.0 in one shared rank range, 2 to 4. A ten-point gap there is a tie. The 104 points between H3 Image and Z-Image Turbo are not.

Why the license sits above the score

A character feed is commercial the day it takes a brand deal, a tip or ad revenue. Three of the four top-scoring models fail that test for most readers. The Qwen license defines Non-Commercial as research or evaluation only and forbids any commercial purpose without a separate license. Fibo 2.0's card says CC BY-NC 4.0 and points commercial users to Bria. MiniMax H3 Image's license, dated 10 November, lists the EU, the UK, the Republic of Korea and the USA as Excluded Territories.

Klein is the one that clears the bar near the top. The FLUX Kommunity License lets a Qualifying User put outputs to commercial use, and the definition is a single line.

huggingface.coOpen the page
Section 1.e of the FLUX Kommunity License on Hugging Face, highlighted: a Qualifying User is you if your and your affiliates' gross annualized revenue is less than US$5,000,000.
The clause the ranking rests on: under US$5M revenue, Klein's outputs are yours to sell (section 2.a). Source: FLUX 3 Image Klein LICENSE.md.

Krea 2 draws its line lower. Section 2.3 of the Krea 2 Community License allows commercial use only below US$1,000,000 in company-wide revenue over a trailing twelve months, and section 4.2 says you must implement reasonable content filter measures in your deployment. A solo creator can live with both, but should know they exist. MOSS Image and Z-Image Turbo are Apache-2.0 and have no such lines at all.

References or a LoRA: how each model keeps the face

A recurring character needs the same face in every post. There are two ways to get it, and the models split cleanly. Klein (10), Qwen 2.1 (10), H3 Image (16) and Fibo 2.0 (4) accept reference photos, so you can pass your character's best three shots with every request. MOSS Image, Krea 2 Turbo and Z-Image Turbo are text to image only, so the face has to live in a character LoRA.

We compared the two routes on Krea 2 in Identity Edit vs a character LoRA; the short version is that references are faster to start and LoRAs drift less over a hundred posts. Our consistent character guide covers the rest of the setup.

Musetta studioDemo data
The Musetta studio home screen with a request typed for the demo character Ines on a rooftop garden at golden hour, 9:16, on FLUX 3 Image Klein distilled at 16 steps and CFG 1 with three reference photos and the prompt upsampler.
Staged with demo data: the ranking's winner asked for by name, with the character's references and Klein's own step count.

What creators are saying

What creators are saying

The pattern is the one the ranking predicts: the license threads are about H3 Image, and the quality threads are about the smaller models. Nobody has published a same-character test across all seven, and neither have we, so the face-keeping order above comes from what each model accepts, not from a measurement.

Which one to pick, by situation

Cheat sheet: one pick per situation

Paid feed, under $5M revenue, 12 to 16 GB card
FLUX 3 Image Klein, distilled, 16 steps, with references
No license ceiling wanted
MOSS Image plus a character LoRA
8 GB card
Z-Image Turbo (8.33 GB kit) or Klein INT8
Living on Civitai LoRAs, under $1M revenue
Krea 2 Turbo, 8 steps
Outside US, EU, UK, KR and best quality first
MiniMax H3 Image, 6 steps
Testing looks only, no posting
Qwen-Image 2.1 or Fibo 2.0

If you already run H3 Image and the license question bites, our H3 Image review and GGUF guide cover the files. For Klein's settings, start with the Klein workflow breakdown: BFL's own docs say to always run the prompt upsampler, which is most of what separates good Klein portraits from the ones in the complaint threads.

How we ranked

  1. License first: can a solo creator with revenue under $1M sell the outputs, anywhere? Read at each model's own LICENSE or card on 17 Nov 2026.
  2. Reference input second: how many photos of the character the model accepts, per its card or launch post.
  3. Score third: Artificial Analysis open-weights text-to-image Elo, read the same morning. Models not on the board get no score.
  4. Weight fourth: the smallest Comfy-Org kit, summed from the Hugging Face file API in decimal GB.
  5. No face test: we did not generate images for this list, and say so wherever it matters.

The one paragraph about us: Musetta runs on your own ComfyUI, installs the nodes and model files a workflow is missing, and keeps each character's reference photos attached to every request, so switching from one model on this list to another does not mean rebuilding the character. The features page shows what it does today.

Questions

What is the best open source image model in 2026?
By score, MiniMax H3 Image (Elo 1066 on the Artificial Analysis open-weights board). For a creator who sells content, FLUX 3 Image Klein, because its outputs are licensed for commercial use under US$5M revenue and it accepts up to 10 reference images.
Can I use Qwen-Image 2.1 for a monetized account?
Not under its open license. The Qwen Research License defines Non-Commercial as research or evaluation only and forbids commercial use without a separate license from Alibaba.
Can I use MiniMax H3 Image in the US or EU?
The license grants no rights in the Excluded Territories: the EU, the UK, the Republic of Korea and the USA. Elsewhere, authorization is needed above US$1M annualized revenue.
Which open image model is fully free for commercial use?
MOSS Image and Z-Image Turbo are Apache-2.0, with no revenue limit or territory list. Klein and Krea 2 allow commercial outputs below $5M and $1M revenue respectively.
Which model runs on 8 GB of VRAM?
Z-Image Turbo has the lightest kit at 8.33 GB, and BFL says Klein's INT8 build runs on RTX cards with as little as 8 GB.

Sources

  1. 01AA-Image-T2I v2.1 open weights leaderboard · Artificial Analysis, 2026-11-17
  2. 02MiniMax-H3-Image Community License Agreement · MiniMax, 2026-11-10
  3. 03Qwen-Image-2.1 license · Alibaba Qwen, 2026-09-20
  4. 04FLUX 3 Image Klein launch post · Black Forest Labs, 2026-10-15
  5. 05FLUX 3 Image Klein LICENSE.md · Black Forest Labs
  6. 06Fibo 2.0 dev model card · Bria
  7. 07Krea 2 Community License Agreement (PDF) · Krea
  8. 08Krea 2 Turbo model card · Krea
  9. 09MOSS-Image-SFT model card · OpenMOSS
  10. 10Z-Image-Turbo model card · Tongyi-MAI
  11. 11Comfy-Org MiniMax-H3-Image files · Comfy-Org
  12. 12FLUX 3 Image Klein ComfyUI files · Black Forest Labs
  13. 13Comfy-Org Krea-2 files · Comfy-Org
  14. 14Comfy-Org MOSS-Image files · Comfy-Org
  15. 15Comfy-Org Qwen-Image-2.1 files · Comfy-Org
  16. 16Comfy-Org Z-Image-Turbo files · Comfy-Org
  17. 17Hugging Face discussion: Q&A about the H3 Image license · Hugging Face community
  18. 18Hugging Face discussion: Klein body types · Hugging Face community

Researched and written by the Musetta editorial desk with AI assistance, from the sources above. Numbers come from those sources; nothing is benchmarked by us unless the article says so. Spotted an error? Tell us on Discord and we fix it.