Best Open-Source AI Image Models, Measured (2026)

Three of every four LoRAs published on Civitai in the past month were made for two anime models. Anima and Illustrious XL drew 16,934 of the 22,959 new LoRAs, 73.8% of the total, while every FLUX model combined drew 325, or 1.4%.

We pulled data from five public sources: every LoRA and checkpoint published on Civitai between 7 September and 7 October through its public API, the Hugging Face Hub API for 29 open image model families (the official repos, the Comfy-Org repackages, and all 66,789 community models that declare one of them as their base in the model tree), the license of every family, and the source code of five creator apps that run these models. Here is what the numbers say about which open image model is worth your time, which ones you can use for paid work, and which ones you can run without touching a terminal.

Key Findings

1. Anime models took three quarters of all new LoRAs

A LoRA is a small add-on file that teaches an image model a style, a character or a look. It is the unit of work for the open image community, and Civitai is where most of them are published. In the past month, 9,239 new LoRAs named Anima as their base and 7,695 named Illustrious, together 73.8% of everything uploaded. Add NoobAI and Pony, two older anime models built on SDXL, and the figure reaches 76.1%.

Anima is the surprise. It is a 2 billion parameter model released this year by CircleStone Labs in collaboration with Comfy Org, built on NVIDIA's Cosmos-Predict2 and trained on several million anime images plus about 800,000 non-anime artworks. Its model card says plainly that it is designed for illustration and "will not work well at realism". Creators did not need it to: it also leads Civitai checkpoints, with 205 of the 562 published in the month (36.5%).

2. Krea 2 is the only general-purpose model with a real community

Outside anime, one model is pulling creators in. Krea 2, which Krea released as open weights in June in two versions (Raw and Turbo), drew 3,813 new LoRAs on Civitai (16.6%) and 143 new checkpoints (25.4%). That is almost ten times the LoRA count of the next general model, SDXL 1.0, at 391.

Hugging Face tells the same story from the other side. Of the 788 new community LoRAs published on Hugging Face for the 29 families we measured, 460 were for Krea 2, 58.4% of the total. Krea 2 already has 2,626 community LoRAs on the Hub, the third largest back catalogue after FLUX.1 dev and SDXL, built in under four months. It also has the friendliest license of the leaders for working creators: free commercial use, images included, for anyone with less than $1 million in annual revenue (more on that in finding 5). We covered the model in Krea 2: foundation image model with built-in style transfer.

3. FLUX has the biggest past and almost no present

FLUX.1 dev remains the largest single ecosystem on Hugging Face: 42,701 community LoRAs declare it as their base, more than every other family we measured combined. But only 51 of them appeared in the past month, and only 214 new FLUX.1 dev LoRAs reached Civitai. Across every FLUX label on Civitai (FLUX.1 dev, schnell and Krea, FLUX.2 dev and all four FLUX.2 klein variants) the total was 325 new LoRAs, 1.4%.

The newer FLUX models are popular to run but not to train. FLUX.2 klein logged 3.06 million downloads in the month and appears in 392 Hugging Face Spaces, more than any other family, yet drew 94 Civitai LoRAs. FLUX.2 dev, a 32 billion parameter model, drew 13.

The Anima row is the clearest sign that the two platforms measure different crowds: 9,239 new Anima LoRAs on Civitai against 23 on Hugging Face. Anyone judging the open image world from Hugging Face alone would miss its biggest community entirely.

4. The most downloaded model is not the one people build on

Z-Image, Alibaba Tongyi's 6 billion parameter model, was the most downloaded family on Hugging Face: 9.61 million downloads in the month, 21.2% of all downloads across the 29 families, most of them through the Comfy-Org repackage that ComfyUI users pull. Yet it drew 242 new Civitai LoRAs, 1.1% of the total.

Qwen-Image-2.1 is the other outlier. Released in mid-September, it logged 7.31 million downloads and 301 new community models on Hugging Face in about three weeks, 23.8% of all new community models we counted, mostly quantized versions (132) and finetunes (60) that shrink it for smaller graphics cards. Downloads tell you what people try. LoRAs tell you what they commit to.

5. A quarter of new work sits on models you cannot use for paid jobs

Open weights do not mean free to use commercially, and the license that matters most for a working creator is the one that covers the images. We read the license of every leading family. The surprise is that two of the "non-commercial" licenses are kinder to creators than they look: the FLUX.1 dev, FLUX.2 and Anima licenses restrict how you run the model (no paid service, no production pipeline) but say outright that you may use the images you generate "for any purpose (including for commercial purposes)".

Two families give no such permission. Qwen-Image-2.1's license defines non-commercial as "research or evaluation purposes only", and the Ideogram 4 agreement covers only use that does not "directly or indirectly generate revenue". Together they account for 303 of the 1,265 new community models on Hugging Face, 24.0%. We explained the Qwen change in Qwen-Image-2.1 trades Apache 2.0 for a research license.

6. 43.6% of new LoRAs say you may not sell what you make with them

The base model's license is only half of the question. Every Civitai LoRA carries its own permissions, set by the person who trained it, and one of them is whether you may sell images generated with it. Of the 22,959 new LoRAs, 10,004 (43.6%) do not allow it. The share varies sharply by base: 72.8% of new Krea 2 LoRAs allow selling images, against 47.9% of Anima LoRAs and 27.0% of NoobAI LoRAs.

For a creator who works for clients, that means two checks, not one: the base model's license, and the small permission icons on each LoRA page. A Krea 2 LoRA that allows selling images, used by a freelancer under the $1 million revenue line, clears both.

7. Every model with a real community runs in at least four creator apps

We checked which families five creator apps can load, from each app's source code: ComfyUI, SwarmUI, InvokeAI, Krita AI Diffusion (the AI plugin for the free painting app Krita) and Draw Things (Mac, iPhone and iPad). Eleven of the 29 families run in all five. Five run in none of them, and those five drew zero Civitai LoRAs between them.

The pattern holds in the other direction too. All seven families with at least 100 new Civitai LoRAs run in at least four of the five apps. The two largest models we measured, HunyuanImage 3.0 at 83 billion parameters and Cosmos3 Text2Image at 65 billion, run in none and drew no Civitai LoRAs. The two biggest LoRA bases, Anima and Illustrious, are both under 3 billion parameters: small enough to train on a home graphics card.

Trend Analysis

Rising: Anima (9,239 new Civitai LoRAs and 205 checkpoints, both the most of any base) and Krea 2 (3,813 LoRAs, 460 of 788 new Hugging Face LoRAs). Qwen-Image-2.1 is rising on downloads (7.31 million) but not yet on LoRAs (120).

Stable: Illustrious XL, which still draws 7,695 LoRAs a month two years after release, and Z-Image, the most downloaded family at 9.61 million.

Declining as a place to build: FLUX. FLUX.1 dev holds the largest back catalogue (42,701 LoRAs) but every FLUX label together drew 1.4% of new Civitai LoRAs.

Not catching on: models that need a data-center GPU or a custom setup. The five families no creator app supports drew 23 new community models on Hugging Face and none on Civitai. Microsoft's own repos for Mage-Flow and Lens no longer load publicly on Hugging Face; the copies that remain are Comfy-Org repackages.

Predictions

Krea 2 stays the only general-purpose model above 10% of new Civitai LoRAs through the end of 2026. It is at 16.6%, and the next general model, SDXL 1.0, is at 1.7%.

Qwen-Image-2.1 will not pass 1,000 Civitai LoRAs by mid-December. Its download count says people want it, but its research-only license gives LoRA makers no way to sell their work or let others sell theirs, and it drew 120 LoRAs in its first weeks.

Every FLUX label combined stays under 2% of new Civitai LoRAs. It is at 1.4%, and the most downloaded new FLUX model, FLUX.2 klein, drew 94.

The next open image model to pass 5% of new Civitai LoRAs will run in at least four of the five creator apps within a month of release. All seven families above 100 LoRAs already do, and none of the five unsupported families has a single Civitai LoRA.

What This Means for Creators

If you draw anime or illustration: Anima is where the new styles and characters are appearing, and it runs in ComfyUI, SwarmUI, InvokeAI and Krita AI Diffusion. Illustrious XL has a fully commercial license and runs everywhere, including Draw Things on a Mac or iPad, which does not support Anima. About four in ten new LoRAs for both carry Civitai's mature-content flag, so set your browsing filter before you search at work.

If you make photos, product shots or brand work: start with Krea 2. It has the community, it runs in all five apps, and its license covers commercial use if your business earns under $1 million a year. Z-Image (Apache 2.0) is the faster, more permissive alternative with fewer LoRAs.

If you need readable text in your images: Ideogram 4 is strong at typography but non-commercial, so it is for personal projects only. For client work, Qwen-Image under Apache 2.0 is the safe open choice, and the older 2512 version is the one to download: version 2.1 is research only. Our text rendering comparison covers the paid options.

If you sell your images: check two things before you rely on a LoRA, the base model's license and the LoRA's own permission icons. 43.6% of new LoRAs do not allow selling images.

If you would rather not touch a terminal: pick from the eleven families that run in all five apps. Draw Things is the simplest start on a Mac, iPhone or iPad; Krita AI Diffusion if you already paint in Krita; InvokeAI if you want a canvas. Every model with a real LoRA community is in at least four of them.

Full Data

All 29 families we measured. Monthly downloads count the official repos plus the Comfy-Org repackage, as the Hugging Face Hub reports them. Community models are repos that declare a family repo as their base (LoRAs, quantized versions, finetunes and merges), one level deep. Apps counts how many of ComfyUI, SwarmUI, InvokeAI, Krita AI Diffusion and Draw Things load the family.

App support was read from each app's source code: ComfyUI supported_models.py, the InvokeAI architecture definitions, the Krita AI Diffusion architecture list, the SwarmUI model class list and the Draw Things model zoo. Illustrious XL loads anywhere SDXL does, since it shares the architecture. Civitai counts include every LoRA whose newest version was published in the window, mature content included, by the base model its creator declared.

Frequently asked questions

What is the best open-source AI image model in 2026?

By community adoption, Krea 2 for general images and Anima for anime and illustration. Krea 2 drew 3,813 new Civitai LoRAs and 460 of 788 new Hugging Face LoRAs in the past month, runs in all five creator apps we checked, and allows commercial use under $1 million in annual revenue. Anima drew 9,239 new Civitai LoRAs, the most of any model.

Which open image model has the most LoRAs?

FLUX.1 dev has the largest back catalogue on Hugging Face, 42,701 community LoRAs. For new work, Anima leads: 9,239 new Civitai LoRAs in the past month, 40.2% of all new LoRAs, followed by Illustrious XL with 7,695 and Krea 2 with 3,813.

Can I sell images made with FLUX.1 dev?

Yes. The FLUX.1 [dev] Non-Commercial License restricts how you use the model itself, but its outputs clause says you may use generated images for any purpose, including commercial purposes, except to train a competing model. The same wording appears in the FLUX.2 license.

Is Qwen-Image-2.1 free for commercial use?

No. Qwen-Image-2.1 ships under a Qwen Research license that defines non-commercial use as research or evaluation only, and commercial use needs a separate license from the Qwen team. The earlier Qwen-Image 1.0, Qwen-Image-2512 and Qwen-Image-Edit models remain under Apache 2.0.

Which open image models run on a Mac or iPad without a terminal?

Draw Things, a Mac, iPhone and iPad app, supports Krea 2, Z-Image, FLUX.2 klein, FLUX.2 dev, Qwen-Image, Qwen-Image-2.1, Ideogram 4, HiDream-O1, SDXL, Illustrious XL, SD 3.5 and ERNIE-Image. It does not support Anima; on a Mac, Anima runs in Krita AI Diffusion or InvokeAI.

What is Anima?

Anima is a 2 billion parameter text-to-image model from CircleStone Labs and Comfy Org, built on NVIDIA's Cosmos-Predict2 and trained mainly on anime images and other non-photorealistic art. It is released under a non-commercial license that still allows commercial use of the images you generate, and it is now the most popular base for new LoRAs on Civitai.

This research was produced by Creative AI News.

Subscribe for free to get the weekly digest every Tuesday.