Best open-weight image generation models
Best open-weight image generation models is a comparison of image generation and image editing models whose trained weights can be downloaded and run on a user's own hardware, current to 19 November 2026. It covers 29 model lines and groups them by the licence attached to the weights (permissive, conditional or non-commercial) and then by documented capability. It does not place them in a single ranked order. Quality evidence comes from two public leaderboards run by Artificial Analysis and from download counts on Hugging Face, each quoted with its date. [1][2][3]
Every model here is called open-weight, not open source. The Open Source AI Definition 1.0 of the Open Source Initiative asks for more than parameters under OSI-approved terms: it also requires the code used to train and run the system and "Data Information", which it defines as "Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system". [4] A permissive licence on a weights file does not by itself meet that definition. Hosted and closed models are covered in Best AI image generators, and video models in Best open-weight video models.
At a glance
The table gives, for each licence tier, the model in this article with the highest rating on each Artificial Analysis leaderboard when the boards were read on 19 November 2026. The ratings are Artificial Analysis's, not AI Wiki's, and they change as votes accumulate. [1][2]
| Licence tier | What the weights licence allows | Highest text-to-image rating (AA-Image-T2I v2.1) | Highest editing rating (AA-Image-Editing v2.1) |
|---|---|---|---|
| Permissive | Commercial use, modification and redistribution, with attribution and notice conditions | Ming-Image-0.1-Design, 1001 ±9 | Qwen-Image-Edit-2511, 996 ±16 |
| Conditional | Commercial use subject to revenue ceilings, use restrictions, territorial limits or attribution | MiniMax-H3-Image, 1066 ±14 | MiniMax-H3-Image, 1135 ±16 |
| Non-commercial | Research or other non-commercial use; commercial use needs a separate licence | Qwen-Image-2.1, 1047 ±12 | Fibo 2.0, 1086 ±13 |
On both boards the three highest-rated models with downloadable weights were under conditional or non-commercial licences. Artificial Analysis's own summary named H3-Image (1066), Qwen-Image-2.1 (1047) and FLUX 3 Image Klein (1041) as the leading open-weights models for text-to-image, and H3-Image (1135), Fibo 2.0 (1086) and Qwen-Image-2.1 (1082) for editing. [1][2] The highest-rated model overall on the text-to-image board that day was a hosted model, GPT Image 2.6 Sunburst (max), at 1261. [1]
How the models were selected
The list is AI Wiki's own selection and is not exhaustive. A model line was considered if it met three criteria on 19 November 2026:
- Its weights are published in a public repository, either openly or behind a click-through gate.
- The licence text that governs those weights, or the developer's licence statement where a repository has no licence file, could be opened and read.
- There is independent evidence of quality or adoption: a row marked as open weights on an Artificial Analysis image leaderboard or, failing that, downloads recorded by the Hugging Face Hub.
The third criterion sets no threshold and was applied by judgement. Seven of the 29 lines have no open-weights row on either Artificial Analysis board and are included on download figures alone. The Hub's figures for their repositories were: SDXL 1.0, 4,794,303; Stable Diffusion 1.5 (mirror), 1,195,331; Krea 2 Raw, 106,131, and Krea 2 Turbo, 83,304; Chroma1-HD, 93,829; Stable Diffusion 3.5 Medium, 76,258, Large, 33,183, and Large Turbo, 3,372; MOSS-Image-SFT, 21,055, and MOSS-Image-Base, 12,339; HiDream-I1-Full, 605. [3] Two of these lines, Stable Diffusion 3.5 Large and Krea 2 Turbo, have rows on the Arena page described below. [5]
Some rows that Artificial Analysis marked as open weights are not covered here, because their licences were not reviewed for this article: OmniGen V2, Bagel, FIBO, FIBO Lite, FIBO Edit, SRPO, Step1X-Edit and HiDream-E1.1, and the FLUX.2 [dev] Turbo and Flash and Ideogram 4.0 Instant and Fast rows, which link to repositories published by a third party. [1][2]
Tables are ordered by release date within each licence tier, unless a table says otherwise. The order is not a quality ranking.
Leaderboards
AA-Image-T2I v2.1 rates models on generating an image from a text prompt, and AA-Image-Editing v2.1 rates them on editing an image from a text instruction. Both use an Elo rating computed from blind pairwise comparisons: Artificial Analysis says the votes come from its "validated private evaluators", who compare two results for the same input without knowing which model made each. [1][2] On 19 November 2026 the text-to-image board had 116 rows and the editing board 86. FLUX.2 [dev] is the anchor of both scales, "Fixed at 1000 as the reference point for this board's Elo scale". [1][2]
Three cautions apply to every figure in this article:
- Ratings move as votes are added and as new models enter, so a figure is valid only for its date.
- Each rating is published with a 95% interval. Two models whose intervals overlap are not separated by the data.
- Ratings from different leaderboards are on different scales and are not comparable. A text-to-image rating cannot be compared with an editing rating, and neither can be compared with a rating from another site.
The Arena site (formerly LMArena) runs a separate text-to-image leaderboard. The live page could not be read in full in the research for this article, so the figures below come from an Internet Archive copy of the page dated 2 November 2026, which listed 84 models and 6,961,993 votes. That page predates MiniMax-H3-Image and had no row for FLUX 3 Image Klein, Fibo 2.0, Ming-Image-0.1-Design, ERNIE-Image or LongCat-Image. The rows for checkpoints covered in this article are listed in the order of the Arena page. [5]
| Model as listed by Arena | Weights licence tier | Arena rank (of 84) | Arena score |
|---|---|---|---|
| Qwen-Image-2.1 | Non-commercial | 21 | 1227 ±6 (marked preliminary) |
| Ideogram 4.0 (Quality) | Non-commercial | 22 | 1199 ±4 |
| Cosmos3-Super-Text2Image (Agentic) | Permissive | 29 | 1165 ±7 |
| Cosmos3-Super-Text2Image | Permissive | 35 | 1151 ±6 |
| HunyuanImage 3.0 | Conditional | 36 | 1151 ±3 |
| FLUX.2 [dev] | Non-commercial | 40 | 1145 ±4 |
| Qwen-Image (2512) | Permissive | 47 | 1125 ±3 |
| HiDream-O1-Image | Permissive | 53 | 1115 ±4 |
| Krea 2 Turbo | Conditional | 54 | 1115 ±4 |
| Z-Image-Turbo | Permissive | 61 | 1082 ±5 |
| FLUX.2 [klein] 9B | Non-commercial | 64 | 1070 ±3 |
| Qwen-Image | Permissive | 68 | 1057 ±3 |
| FLUX.2 [klein] 4B | Permissive | 72 | 1030 ±3 |
| GLM-Image | Permissive | 78 | 1012 ±9 |
| FLUX.1 [dev] (FP8) | Non-commercial | 80 | 969 ±4 |
| FLUX.1 Kontext [dev] | Non-commercial | 82 | 941 ±4 |
| Stable Diffusion 3.5 Large | Conditional | 83 | 938 ±4 |
Arena scores are on Arena's own scale. They are not combined with the Artificial Analysis figures, and the tier column is this article's classification, not Arena's.
Download counts
The Hugging Face Hub shows a "Downloads last month" figure for each repository, and its API returns the same number. Hugging Face's documentation says the count is taken from requests for designated files in a repository and that "Every HTTP request to these files, including GET and HEAD, will be counted as a download". [71] The figures below were read on 19 November 2026 for a selection of repositories of the model lines in this article. They are listed from highest to lowest, and the table is not a complete list of repositories. [3]
| Repository | Downloads last month (Hub figure) | Weights licence tier |
|---|---|---|
| stabilityai/stable-diffusion-xl-base-1.0 | 4,794,303 | Conditional |
| stable-diffusion-v1-5/stable-diffusion-v1-5 | 1,195,331 | Conditional |
| black-forest-labs/flux-3-image-klein-comfy | 1,037,285 | Conditional |
| Tongyi-MAI/Z-Image-Turbo | 807,620 | Permissive |
| MiniMaxAI/MiniMax-H3-Image | 759,661 | Conditional |
| black-forest-labs/FLUX.2-klein-base-4B | 710,603 | Permissive |
| Qwen/Qwen-Image-Edit-2509 | 605,537 | Permissive |
| stabilityai/sdxl-turbo | 585,974 | Conditional |
| black-forest-labs/FLUX.2-klein-4b-fp8 | 565,174 | Permissive |
| black-forest-labs/FLUX.1-dev | 451,324 | Non-commercial |
| black-forest-labs/FLUX.2-klein-4B | 437,305 | Permissive |
| Qwen/Qwen-Image | 373,868 | Permissive |
| black-forest-labs/FLUX.2-dev | 288,741 | Non-commercial |
| Qwen/Qwen-Image-2.1 | 288,581 | Non-commercial |
| black-forest-labs/FLUX.1-schnell | 287,040 | Permissive |
Per-repository counts understate use, because the same weights are redistributed in repackaged form. On the same date the Comfy-Org copies prepared for ComfyUI recorded 9,550,345 downloads for Qwen-Image-2.1, 4,510,397 for Krea 2 and 2,950,724 for MiniMax-H3-Image, several times the figures for the developers' own repositories. [3] The counts measure file requests, not image quality, and the two earliest model generations in this article, SDXL 1.0 and Stable Diffusion 1.5, had the most downloads of any developer or mirror repository listed while having no row on the Artificial Analysis text-to-image board. [3][1]
Repositories for FLUX.1 [schnell], FLUX.1 [dev], FLUX.2 [dev], FLUX.2 [klein] 9B, FLUX 3 Image Klein, Stable Diffusion 3.5, Krea 2, Ideogram 4 and Fibo 2.0 were gated on that date: the files can be fetched only from a Hugging Face account that has accepted the publisher's conditions. The SDXL, Qwen, Z-Image, FLUX.2 [klein] 4B, HunyuanImage 3.0 and MiniMax-H3-Image repositories were not gated. [3]
Licence tiers
The three tiers follow the terms of the weights licence only. Inference code is often under a different licence from the weights, and text encoders or prompt tools bundled with a model can carry their own terms. For the general background see Open-source license and Open-weight license comparison, which covers language-model licences in more detail.
In the tables, "(repo)" after a date means the developer's card gives no release date and the date is the creation date of the Hugging Face repository. [3] The rating column is the AA-Image-T2I v2.1 Elo read on 19 November 2026; "no row" means the model was not listed as an open-weights row on that board. [1]
Permissive: Apache 2.0, MIT and OpenMDW
These weights are under the Apache License 2.0, the MIT License or OpenMDW 1.1. None sets a revenue ceiling, a field-of-use restriction or a territorial exclusion. OpenMDW-1.1 grants permission "to deal in the Model Materials without restriction" and places no restriction on outputs. [6]
| Model | Developer | Weights released | Parameters | Weights licence | AA text-to-image Elo |
|---|---|---|---|---|---|
| FLUX.1 [schnell] | Black Forest Labs | 2024-08-01 | 12 billion | Apache 2.0 | 743 ±19 |
| HiDream-I1 | HiDream-ai | 2025-04-06 (repo) | 17 billion | MIT (transformer) | no row |
| Qwen-Image and Qwen-Image-2512 | Alibaba (Qwen) | 2025-08-04; 2025-12-31 | 20 billion | Apache 2.0 | 908 ±16 |
| Qwen-Image-Edit, Edit-2509, Edit-2511 and Qwen-Image-Layered | Alibaba (Qwen) | 2025-08-18 to 2025-12-23 | built on Qwen-Image | Apache 2.0 | editing board only |
| Chroma1-HD (based on FLUX.1 [schnell]) | lodestones | 2025-08-08 (repo) | 8.9 billion | Apache 2.0 | no row |
| Z-Image-Turbo and Z-Image | Tongyi-MAI | 2025-11-26; 2026-01-27 | 6 billion | Apache 2.0 | 937 ±11 (Turbo) |
| LongCat-Image | Meituan | 2025-12-04 (repo) | 6 billion | Apache 2.0 | 862 ±18 |
| GLM-Image | Z.ai | 2026-01-08 (repo) | 9 billion autoregressive plus 7 billion decoder | MIT | 890 ±16 |
| FLUX.2 [klein] 4B | Black Forest Labs | 2026-01-15 | 4 billion | Apache 2.0 | 840 ±12 |
| ERNIE-Image and ERNIE-Image-Turbo | Baidu | 2026-04-07 (repo); 2026-04-02 (repo) | 8 billion | Apache 2.0 | 909 ±10; 915 ±10 |
| HiDream-O1-Image and O1-Image-Dev | HiDream-ai | 2026-05-08 | 8 billion | MIT | 979 ±10; 810 ±19 |
| Cosmos3-Super-Text2Image | NVIDIA | 2026-05-31 | 64 billion | OpenMDW-1.1 | 988 ±10 |
| Ming-Image-0.1-Design | inclusionAI | 2026-09-17 (repo) | 6 billion | MIT | 1001 ±9 |
| MOSS-Image-Base and MOSS-Image-SFT | OpenMOSS Team | 2026-10-16 | 5 billion | Apache 2.0 | no row |
Sources for the rows: FLUX.1 [schnell][7][8]; HiDream-I1 [9]; Qwen-Image and its editing models [10][11][12]; Chroma1-HD [13]; Z-Image [14][15]; LongCat-Image [16]; GLM-Image [17]; FLUX.2 [klein] 4B [18][19]; ERNIE-Image [20]; HiDream-O1-Image [21]; Cosmos3-Super-Text2Image [22]; Ming-Image-0.1-Design [23]; MOSS-Image [24].
Notes on individual rows:
- FLUX.1 [schnell] is a distilled 12 billion parameter rectified flow transformer, and its card says it can be used "for personal, scientific, and commercial purposes". The repository is gated although the licence is Apache 2.0. [8][3]
- HiDream-I1's MIT licence covers the transformer. The card states that the VAE comes from FLUX.1 [schnell] (Apache 2.0) and that the text encoders are T5 (Apache 2.0) and Meta's Llama 3.1 8B Instruct, which is under the Llama 3.1 Community License. The same section of the card says "Do not create illegal content, harmful material, personal information that could harm others, false information, or content targeting vulnerable groups"; that instruction is the card's and is not a term of the MIT licence. [9]
- HiDream-O1-Image is a pixel-level transformer with no separate VAE or text encoder. Its optional prompt agent can use Gemma-4-31B-it, and the card notes that this requires accepting the Gemma licence. [21]
- Artificial Analysis also lists a row named "Z-Image Base" (894 ±16), which it did not mark as open weights on that date, so the figure is not used in the table. [1]
- GLM-Image combines an autoregressive generator with a diffusion decoder. Its card says the VQ tokenizer and ViT weights come from X-Omni under Apache 2.0. [17]
- Cosmos3-Super-Text2Image is described by NVIDIA as "ready for commercial and non-commercial use". On the Artificial Analysis board it also appears in an agentic configuration (995) and a 4-step configuration (973). [22][1]
- Black Forest Labs described FLUX.2 [klein] as an "Open-source, Apache 2.0 model" when it announced the FLUX.2 family; that is the developer's wording. Only the 4B models are under Apache 2.0. [25][18]
Conditional: commercial use with conditions
These licences allow some commercial use but attach conditions that the permissive licences do not.
| Model | Developer | Weights released | Parameters | Weights licence | Main conditions | AA text-to-image Elo |
|---|---|---|---|---|---|---|
| SDXL 1.0 | Stability AI | 2023-07-26 | 3.5 billion (base) | CreativeML Open RAIL++-M | Use-based restrictions that must be passed on to recipients | no row |
| Stable Diffusion 1.5 | community mirror of a deprecated Runway repository | 2024-08-30 (mirror repo) | not stated on card | CreativeML OpenRAIL-M | Use-based restrictions | no row |
| Stable Diffusion 3.5 Large, Large Turbo and Medium | Stability AI | 2024-10-22; Medium 2024-10-29 | 8.1 billion; Medium 2.5 billion | Stability AI Community License | Commercial use only below USD 1 million annual revenue, with registration | no row |
| HunyuanImage 3.0, Instruct and Instruct-Distil | Tencent | 2025-09-28; Instruct 2026-01-26 | 80 billion (13 billion active per token) | Tencent Hunyuan Community License | Does not apply in the EU, UK or South Korea; separate licence for licensees above 100 million monthly active users | 933 ±15; Instruct 986 ±11 |
| Krea 2 Raw and Krea 2 Turbo | Krea.ai, Inc. | 2026-06-22 | 12 billion | Krea 2 Community License v.1 | Commercial use only below USD 1 million revenue over the trailing twelve months | no open-weights row |
| FLUX 3 Image Klein (distilled and Base) | Black Forest Labs | 2026-10-15 | 5.6 billion | FLUX Kommunity License v.1.0 | Commercial use of outputs only below USD 5 million gross annualized revenue | 1041 ±11 |
| MiniMax-H3-Image | MiniMax | 2026-11-10 | 33.1 billion | MiniMax-H3-Image Community License | Excludes the EU, UK, South Korea and the US; written authorization above USD 1 million annualized revenue or for hosted services | 1066 ±14 |
Sources for the rows: SDXL 1.0 [27][28]; Stable Diffusion 1.5 [26]; Stable Diffusion 3.5 [29][30]; HunyuanImage 3.0 [31][32][33]; Krea 2 [34][35]; FLUX 3 Image Klein [36][37][38]; MiniMax-H3-Image [39][40].
What the conditions say, licence by licence:
Stable Diffusion 1.5 and SDXL 1.0 use OpenRAIL licences. The CreativeML Open RAIL++-M licence dated July 26, 2023, which covers SDXL 1.0, permits hosting for third-party remote access and redistribution, and lists prohibited uses in an attachment (Attachment A) that must be passed on with the model and its derivatives. The licensor claims no rights in the output. [28]
The Stability AI Community License, which covers Stable Diffusion 3.5 and was last updated July 5, 2024, requires commercial users to register and ends the licence when the licensee or its affiliates generate more than USD 1,000,000 in annual revenue. It requires a "Powered by Stability AI" notice and forbids using the models or outputs "to create or improve any foundational generative AI model". [30]
The Tencent Hunyuan Community License states that it "DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA". A licensee whose products or services had more than 100 million monthly active users in the calendar month before the release date must request a licence from Tencent, and the licence forbids using the model or its outputs "to improve any other AI model". The licence file is in the HunyuanImage-3.0 and HunyuanImage-3.0-Instruct repositories. The HunyuanImage-3.0-Instruct-Distil repository had no licence file on 19 November 2026, and its Hub licence tag read "other" with no licence named. [32][3]
Under the Krea 2 Community License v.1, commercial use is allowed only for a licensee whose company-wide revenue, affiliates included, is below USD 1,000,000 on a trailing twelve-month basis; others need an enterprise licence. Distributed derivative models must carry names beginning with "Krea", and content filters are required. [35] The parameter count differs by source: the model card says 12 billion, the Hub's file metadata for each of the two repositories totals 12.82 billion (displayed as 13B), and Black Forest Labs gives 12.9 billion for Krea 2 in its FLUX 3 Image Klein announcement. [34][3][36] Artificial Analysis lists hosted Krea 2 models (Krea 2 Large at 1035, Krea 2 Medium Turbo at 1020 and Krea 2 Medium at 1014) but did not mark them as open weights, so no rating is assigned to the Raw and Turbo checkpoints here. [1]
The FLUX Kommunity License v.1.0 defines a "Qualifying User" as one whose gross annualized revenue, affiliates included, is below USD 5,000,000. Such a user may use outputs commercially, with content filtering or review. The model card adds that the "commercial-use exception is not blanket permission for commercial fine-tuning or distribution". The licence bans military, surveillance and biometric uses. [38][37]
The MiniMax-H3-Image Community License, dated November 10, 2026, applies worldwide except in the "Excluded Territories", defined as "the European Union, the United Kingdom, the Republic of Korea and the United States of America". Prior written authorization is required when the annualized revenue of the licensee and its affiliates exceeds USD 1 million, or when the licensee offers a hosted service to third parties. Commercial products must display the model name. [40]
Non-commercial: research and evaluation licences
These weights may be downloaded and run, but not used commercially without a separate agreement with the developer.
| Model | Developer | Weights released | Parameters | Weights licence | AA text-to-image Elo |
|---|---|---|---|---|---|
| FLUX.1 [dev] | Black Forest Labs | 2024-08-01 | 12 billion | FLUX.1 [dev] Non-Commercial License v1.1.1 | 816 ±33 |
| FLUX.1 Kontext [dev] | Black Forest Labs | 2025-06-26 | 12 billion | FLUX.1 [dev] Non-Commercial License v1.1.1 | editing board only |
| FLUX.1 Krea [dev] | Black Forest Labs with Krea AI | 2025-07-31 | 12 billion | FLUX.1 [dev] Non-Commercial License v1.1.2 | 870 ±29 |
| FLUX.2 [dev] | Black Forest Labs | 2025-11-25 | 32 billion | FLUX Non-Commercial License v2.1 | 1000 (anchor) |
| FLUX.2 [klein] 9B | Black Forest Labs | 2026-01-15 | 9 billion | FLUX Non-Commercial License v2.1 | 931 ±10 |
| Ideogram 4 | Ideogram | 2026-06-03 | 9.3 billion | Ideogram Non-Commercial Model Agreement | 1003 ±10; Quality setting 1026 ±10 |
| Qwen-Image-2.1 and Qwen-Image-2.1-Turbo | Alibaba (Qwen) | 2026-09-20; 2026-10-09 | 7 billion | Qwen Research License | 1047 ±12 |
| Fibo 2.0 | Bria | 2026-10-18 (repo) | 8 billion | CC BY-NC 4.0 (stated on the card) | 1037 ±11 |
Sources for the rows: FLUX.1 [dev][7][41][42]; FLUX.1 Kontext [dev][43][44]; FLUX.1 Krea [dev][45][46]; FLUX.2 [dev][25][47][48]; FLUX.2 [klein] 9B [18][49]; Ideogram 4 [50][51]; Qwen-Image-2.1 [52][53][54]; Fibo 2.0 [55].
The non-commercial licences differ on what may be done with generated images:
- The FLUX.1 [dev] licence (version 1.1.1 in the FLUX.1 [dev] repository) says "You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein". The restriction falls on running the model commercially and on using outputs to train a competing model. [42] The FLUX Non-Commercial License v2.1 keeps both points. [48] The licence files differ by repository: the FLUX.1 Kontext [dev] repository also carries version 1.1.1, the FLUX.1 Krea [dev] repository carries version 1.1.2, and the FLUX.2 [dev] and FLUX.2 [klein] 9B repositories carry version 2.1. [44][46][48][49] The licence page on the Black Forest Labs website was headed "FLUX [dev] Non-Commercial License v2.0", last updated November 25, 2025, when it was read on 19 November 2026. [72]
- The Ideogram agreement, last updated June 3, 2026, excludes from non-commercial use the generation of output to include in, advertise or promote revenue-generating products or services. [51]
- The Qwen Research License, dated September 20, 2026, limits use to "NON-COMMERCIAL PURPOSES ONLY" and defines that as research or evaluation. [54]
- Fibo 2.0 is under Creative Commons Attribution-NonCommercial 4.0 (see Creative Commons licenses), and its card says commercial use requires a commercial agreement with Bria. The repository has no separate licence file; its Hub licence tag is "bria-fibo" and links to the Creative Commons deed. [55][3]
- Artificial Analysis also lists Qwen-Image-2.1-Turbo (1025 ±12) without marking it as open weights, so that figure is not used in the table. [1]
Two model cards in this tier and the conditional tier use the term "open-source" for weights under restricted terms: the Qwen-Image-2.1 card and the MiniMax-H3-Image card. Bria's access notice describes Fibo 2.0 as "open source for non commercial use only". [52][39][55] Those are the developers' descriptions.
Capabilities documented by developers
The groupings below repeat what each developer's card or announcement documents. They are not test results from AI Wiki, and the Artificial Analysis boards measure overall voter preference, not any single capability.
Text rendering and layout
| Model | Licence tier | What the developer documents |
|---|---|---|
| Qwen-Image | Permissive | Complex text rendering, with particular strength in Chinese [11] |
| Z-Image-Turbo | Permissive | Bilingual text rendering in English and Chinese [14] |
| GLM-Image | Permissive | A Glyph Encoder in the decoder for text rendering [17] |
| ERNIE-Image | Permissive | Dense, long-form and layout-sensitive text for posters and infographics [20] |
| Ming-Image-0.1-Design | Permissive | UI, infographics, posters and other text-rich designs, with RGBA output [23] |
| MOSS-Image | Permissive | Layout control through JSON prompts [24] |
| FLUX 3 Image Klein | Conditional | Text rendering and bounding-box layout control [37] |
| Ideogram 4 | Non-commercial | Structured JSON prompting, bounding-box layout and colour-palette controls [50] |
| Qwen-Image-2.1 | Non-commercial | Transparent (RGBA) image generation and layer editing [53] |
Image editing
The rating column is the AA-Image-Editing v2.1 Elo read on 19 November 2026. Rows are grouped by licence tier. [2]
| Model | Licence tier | Editing capability documented | AA editing Elo |
|---|---|---|---|
| Qwen-Image-Edit-2511 | Permissive | Instruction-based editing built on Qwen-Image [12][10] | 996 ±16 |
| HiDream-O1-Image | Permissive | Instruction editing and subject personalization [21] | 952 ±13 |
| FLUX.2 [klein] 4B | Permissive | Text-to-image and multi-reference editing [19] | 941 ±15 |
| MiniMax-H3-Image | Conditional | Editing with up to 16 reference images [39] | 1135 ±16 |
| FLUX 3 Image Klein | Conditional | Editing with up to 10 reference images [37] | 1069 ±13 |
| HunyuanImage 3.0 Instruct | Conditional | Image-to-image editing and multi-image fusion [33] | 1064 ±13 |
| Fibo 2.0 | Non-commercial | Generation and editing in one model [55] | 1086 ±13 |
| Qwen-Image-2.1 | Non-commercial | Editing with up to 10 reference images [52] | 1082 ±13 |
| FLUX.2 [klein] 9B | Non-commercial | Text-to-image and multi-reference editing [49] | 1013 ±12 |
| FLUX.1 Kontext [dev] | Non-commercial | Edits images from text instructions [44] | 883 ±19 |
Artificial Analysis lists the Qwen editing model as "Qwen Image Edit Plus 2511". [2] Z-Image-Edit was announced by its developers but was still marked as to be released on the Z-Image-Turbo card. [14]
Few-step variants
Several developers publish distilled checkpoints that need fewer sampling steps than the base model (see Knowledge distillation). The step counts are those the developers document.
| Model | Licence tier | Documented steps |
|---|---|---|
| FLUX.1 [schnell] | Permissive | 1 to 4 [8] |
| FLUX.2 [klein] 4B | Permissive | 4 [19] |
| Z-Image-Turbo | Permissive | 8 function evaluations [14] |
| ERNIE-Image-Turbo | Permissive | 8, against 50 for ERNIE-Image [20] |
| HiDream-O1-Image-Dev | Permissive | 28, against 50 for the full model [21] |
| SDXL Turbo | Conditional | 1 to 4 [56] |
| Stable Diffusion 3.5 Large Turbo | Conditional | 4 [29] |
| HunyuanImage 3.0 Instruct-Distil | Conditional | 8 [33] |
| Krea 2 Turbo | Conditional | 8 [34] |
| FLUX 3 Image Klein | Conditional | 16, against 50 for the Base checkpoint [37] |
| MiniMax-H3-Image | Conditional | 6, with no classifier-free guidance [39] |
| FLUX.2 [klein] 9B | Non-commercial | 4 [49] |
| Qwen-Image-2.1-Turbo | Non-commercial | 8 [53] |
Memory requirements
Only figures stated by a developer are given, with the conditions the developer attaches. They are not comparable with each other, since some refer to weight file sizes, some to peak GPU memory and some to a recommended card, and several exclude text encoders.
| Model | Stated requirement | Conditions |
|---|---|---|
| SDXL 1.0 | "should work effectively on consumer GPUs with 8GB VRAM" | Stability AI announcement [27] |
| Stable Diffusion 3.5 Medium | 9.9 GB of VRAM | Excluding text encoders [29] |
| FLUX 3 Image Klein | Generator weights of 11.2 GB in BF16 and 5.6 GB in FP8; INT8 build on GPUs "with as little as 8GB of VRAM" | Announcement, through ComfyUI or Diffusers [36] |
| FLUX 3 Image Klein | Peak GPU memory of 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4 | Reference code on a B200 at 1024x1024 with the prompt upsampler off [37] |
| FLUX.2 [klein] 4B | About 13 GB of VRAM | Developer card and announcement [19][18] |
| Z-Image-Turbo | 16 GB of VRAM | "consumer devices", per the card [14] |
| LongCat-Image | About 17 GB | With CPU offload enabled [16] |
| GLM-Image | About 23 GB of GPU memory | With CPU offload, at slower speed [17] |
| ERNIE-Image | 24 GB of VRAM | "consumer GPUs", per the card [20] |
| FLUX.2 [klein] 9B | About 29 GB of VRAM | Developer card [49] |
| HunyuanImage 3.0 | At least three 80 GB GPUs; at least eight for Instruct | "Recommended VRAM" in Tencent's table [33] |
| MiniMax-H3-Image | 66.28 GB transformer file in BF16 plus a 2.34 GB VAE; the text encoder is loaded from the separate MiniMax-H3 repository | File sizes, not a memory measurement [39] |
Black Forest Labs says its FP8 builds of FLUX.2 [klein] use up to 40% less VRAM and its NVFP4 builds up to 55% less. [18] Quantization of this kind trades numerical precision for memory, and the figures above apply only to the precision named. Ideogram publishes Ideogram 4 as quantized nf4 and fp8 weights. [50]
Software for running the models
Weights need an inference runtime, and the runtime has its own licence. Licence and star counts below are from the GitHub API as of 19 November 2026. [57]
| Software | Licence | GitHub stars | What it is |
|---|---|---|---|
| ComfyUI | GPL-3.0 | 147,518 | Node-graph interface and backend; its README lists support for SD 1.5, SDXL, SD 3.5, FLUX.1, FLUX.2, Qwen Image, Z-Image, HiDream, Chroma, LongCat Image, Ideogram 4, Krea 2, ERNIE Image and MOSS Image [58] |
| Diffusers | Apache-2.0 | 34,941 | Python library from Hugging Face [57] |
| stable-diffusion.cpp | MIT | 8,068 | C/C++ inference built on ggml, with GGUF files and Vulkan and Metal backends [59] |
| InvokeAI | Apache-2.0 | 29,117 | Web interface [57] |
| AUTOMATIC1111 Stable Diffusion web UI | AGPL-3.0 | 165,338 | Web interface; last pushed to on 2026-03-02 [57] |
| SGLang | Apache-2.0 | 39,041 | Serving framework with a diffusion backend [57] |
| vLLM-Omni | Apache-2.0 | 7,836 | Serving framework [57] |
ComfyUI's README says it "Can run even the biggest open source models on as low as 4GB vram + 8GB ram" by streaming weights, and that it runs fully offline. [58] That is the project's claim about the software and is not a per-model measurement. The stable-diffusion.cpp README lists SD 3.5, FLUX.1, FLUX.2, FLUX.3 Image Klein, Chroma, Qwen Image, Qwen Image 2.1, Z-Image and Krea 2 among supported models. [59] Black Forest Labs, Qwen and MiniMax each name ComfyUI, Diffusers, SGLang and vLLM-Omni as supported runtimes for their 2026 releases. [36][53][39]
The licence of a model's inference code does not extend to its weights. The code repositories for FLUX 3 Image Klein, Ideogram 4 and Krea 2 are tagged Apache-2.0 on GitHub, while the weights they load are under the restricted licences described above. [57] General guidance on running models on local hardware is in Local AI.
Licence changes and withdrawn repositories
A model's terms are those in force for the checkpoint and date in question. The record below shows how they have moved; a fuller chronology is in Timeline of open-weight models.
| Date | Model or licence | Change |
|---|---|---|
| 2024-07-05 | Stability AI models | Stability AI introduced its Community License, writing that "the commercial license originally associated with SD3 caused some confusion and concern in the community" [60] |
| 2024-07-10 | SDXL Turbo | A commit of 2023-11-28 had uploaded a "Stability Noncommercial Research License" file. On 2024-07-10 the licence file was renamed, and it now contains the Stability AI Community License. The Hub metadata tag still reads sai-nc-community [56] |
| 2024-08-30 | Stable Diffusion 1.5 | A mirror repository was created; its card calls itself a "mirror of the now deprecated" Runway repository and says it is not affiliated with RunwayML. Runway's Hugging Face page reads "We are no longer maintaining a HuggingFace organization" [26][61][3] |
| 2025-06-26 | FLUX.1 [dev] licence | Black Forest Labs updated the licence alongside the Kontext [dev] release, adding a requirement for content filters or manual review, and opened self-serve commercial licensing [43] |
| 2026-01-15 | FLUX licence | The "FLUX [dev] Non-Commercial License" was renamed "FLUX Non-Commercial License" and applied to the 9B Klein models. Black Forest Labs wrote: "No material changes have been made to the license." [18] |
| 2026-07-31 | Sana 1600M | The Efficient-Large-Model/Sana_1600M_1024px repository, whose card names NVIDIA as a developer and which was previously tagged with the NVIDIA Open Model License, was relicensed under Apache-2.0 [62] |
| 2026-09-20 | Qwen-Image family | Qwen-Image-2.1 was published under the Qwen Research License. The earlier Qwen-Image checkpoints remain under Apache 2.0, so the change applies to the new checkpoint and is not a relicensing of the old ones [54][11] |
| 2026-10-15 | FLUX family | FLUX 3 Image Klein was released under the revenue-based FLUX Kommunity License v.1.0, a third set of terms for FLUX image weights beside Apache 2.0 (FLUX.2 [klein] 4B) and the non-commercial licence (FLUX.2 [klein] 9B). The same licence is in the repositories of FLUX 3 Action, a robotics model, which were created on 22 September 2026 [38][18][73][3] |
Several of the conditional and non-commercial licences can be ended by the licensor. The FLUX Kommunity License and the FLUX Non-Commercial License v2.1 both say "We may terminate this License, in whole or in part, at any time upon notice". [38][48] The Ideogram agreement contains a similar clause, and the Krea 2 card says Krea "reserves the right to update model weights or revoke access in response to identified misuse patterns". [51][34] The Krea 2 Community License also lets Krea terminate "for any reason upon thirty (30) days' prior written or electronic notice". [35] The Stability AI Community License ends when the licensee passes its revenue ceiling, as described above. [30]
Limits and risks
Restrictions on outputs and on training other models
Several licences restrict using the model or its images to build other models:
- Stability AI Community License: no use of the models or outputs "to create or improve any foundational generative AI model". [30]
- Tencent Hunyuan Community License: no use of the model or outputs "to improve any other AI model". [32]
- MiniMax-H3-Image Community License and FLUX Kommunity License: both use the phrase "to improve any other artificial intelligence model"; the FLUX clause is limited to models with a function similar to a FLUX model. [40][38]
- FLUX and Ideogram non-commercial licences: outputs may not be used to train, fine-tune or distill a competing model. [42][48][51]
- Qwen Research License: a distributed model trained or improved with the materials or their outputs must display "Built with Qwen" or "Improved using Qwen". [54]
Safety filters and content policies
What ships with the weights varies widely:
- The Stable Diffusion 1.5 card says the intended use is with the Safety Checker in Diffusers, and warns that the training dataset "contains adult material and is not fit for product use without additional safety mechanisms and considerations". [26]
- The FLUX.2 [dev] card says the repository "includes filters for NSFW and IP-infringing content at input and output" and that filters or manual review must be used under the licence. It also says Black Forest Labs may approach known deployers to verify that this is done, and that the company worked with the Internet Watch Foundation to filter known child sexual abuse material from training data. [47]
- The FLUX.1 Kontext [dev] reference code includes an integrity checker, and the FLUX 3 Image Klein reference command-line tool always loads an NSFW image classifier. [44][37]
- Ideogram 4's command-line tool screens content through a third-party service, Hive, and needs that service's API keys. [50]
- The Cosmos3-Super-Text2Image card documents a guardrail option that uses a separate package, cosmos-guardrail. [22]
- The Chroma1-HD card states that the model "has not been aligned with a specific safety filter". [13]
- The Krea 2 and FLUX Kommunity licences make content filters or output review a condition of the licence itself. [35][38]
Use-based restrictions in the OpenRAIL licences and the acceptable-use exhibits of the Tencent and MiniMax licences bind the user even where no filter is shipped. [28][32][40]
Training data and litigation
Developers disclose training data unevenly. The Stable Diffusion 1.5 card says version 1 was trained on subsets of LAION-2B(en). [26] The Stable Diffusion 3.5 cards refer to "synthetic data and filtered publicly available data", Krea says Krea 2 used public, licensed and synthetic data, and Bria says Fibo 2.0 was "Trained exclusively on licensed data". [63][34][55] These statements are the developers' own.
LAION republished its LAION-5B dataset as Re-LAION-5B on 30 August 2024 after the Stanford Internet Observatory reported in December 2023 that it contained links to suspected child sexual abuse material. LAION said 2,236 links were removed. [64]
Researchers have shown that diffusion models can reproduce training images. A 2023 paper by Nicholas Carlini and colleagues reported that diffusion models memorize individual training images, and that the authors extracted over a thousand training examples from the models they tested. [65]
Two court cases concern Stable Diffusion directly. The first is Getty Images v Stability AI in England and Wales. In a High Court judgment handed down on 4 November 2025, Mrs Justice Joanna Smith dismissed Getty Images' claim of secondary copyright infringement, holding that "an AI model such as Stable Diffusion which does not store or reproduce any Copyright Works (and has never done so) is not an 'infringing copy'". She found trade mark infringement in part, describing those findings as "both historic and extremely limited in scope". [66] Getty Images had abandoned its primary copyright and database right claims at trial, on the basis that the training took place outside the jurisdiction. It appealed the dismissal of the secondary infringement claim to the Court of Appeal (reference CA-2026-000007). Its skeleton argument, which says that the trial judge granted permission to appeal, was filed on 31 July 2026, and a replacement was filed on 21 October 2026. [67] No Court of Appeal judgment was found in the research for this article, read on 19 November 2026, and none is reported here. The case has its own article, Getty Images v Stability AI.
The second is Andersen v. Stability AI in the Northern District of California (No. 3:23-cv-00201), a putative class action by artists against Stability AI, Midjourney, DeviantArt and Runway. On 12 August 2024 Judge William H. Orrick denied the motions to dismiss the copyright infringement claims and dismissed Digital Millennium Copyright Act claims with prejudice. [68] A scheduling order of 2 February 2026 set the hearing on class certification for 17 February 2027 and trial for 1 June 2027. [69] The schedule has since been changed: an order of 28 September 2026 set the hearing on class certification and summary judgment for 29 September 2027 and left the dates of the pretrial conference and trial "To be set by Court". [70] The case had not been decided on the merits as of that order.
Neither case has so far produced a ruling on whether training an image model on copyrighted images is lawful. Both concern Stable Diffusion, not the other model families in this article.
Continue reading
- Best AI Image GeneratorsAs of July 2026, the best all-around AI image generator is GPT Image 2 from OpenAI.Listed in See also
- Open-Weight LLM License ComparisonAs of November 2026, the most widely used licenses that put no user, revenue or use conditions on commercial use of open-weight LLMs are two standard, OSI-approved ones: Apache-2.0 (Alibaba's Qwen3, most of…Listed in See also
- List of open-weight modelsThis is a list of open-weight models: artificial intelligence models whose trained parameters (weights) the developer has published for download, so that the model can be run, studied and in most cases…Listed in See also
- Best Open-Source LLMsAs of November 17, 2026, the highest-scoring open-weight large language model is Moonshot AI's Kimi K3.1 Pro, whose weights Moonshot released on November 4.Listed in See also
- Best Local and On-Device LLMsAs of November 2026, the best local LLM to run offline depends almost entirely on how much memory you can give it, so this guide ranks picks by hardware tier.Listed in See also
- Text-to-Image ModelsText-to-image models are generative artificial intelligence systems that synthesize a new image from a natural-language description, called a prompt.Listed in See also
More from See also
More to read
Topics:AI ModelsDiffusion ModelsImage GenerationOpen Source AI
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Artificial Analysis. "AA-Image-T2I v2.1 Leaderboard - Top AI Image Models". Artificial Analysis. Read 19 November 2026. artificialanalysis.ai/...text-to-image
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Artificial Analysis. "AA-Image-Editing v2.1 Leaderboard - Top AI Image Models". Artificial Analysis. Read 19 November 2026. artificialanalysis.ai/...editing
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Hugging Face. Hub API model records ("Downloads last month" figure, gating, licence tag, file list and creation date) for the repositories named in the text. Read 19 November 2026. huggingface.co/...stable-diffusion-xl-base-1.0
- ^Open Source Initiative. "The Open Source AI Definition - 1.0". opensource.org. Read 19 November 2026. opensource.org/...open-source-ai-definition
- ^1 ^2Arena. "Text-to-Image Arena" leaderboard, page dated 2 November 2026. arena.ai. Internet Archive capture of 6 November 2026, read 19 November 2026. web.archive.org/...text-to-image
- ^The Linux Foundation. "OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)". OpenMDW. Read 19 November 2026. openmdw.ai/...1-1
- ^1 ^2Black Forest Labs. "Announcing Black Forest Labs". Black Forest Labs, 1 August 2024. bfl.ai/announcing-black-forest-labs
- ^1 ^2 ^3Black Forest Labs. "black-forest-labs/FLUX.1-schnell" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.1-schnell
- ^1 ^2HiDream-ai. "HiDream-ai/HiDream-I1-Full" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...HiDream-I1-Full
- ^1 ^2QwenLM. "Qwen-Image" (repository README). GitHub. Read 19 November 2026. github.com/...Qwen-Image
- ^1 ^2 ^3Qwen. "Qwen/Qwen-Image" (model card and LICENSE). Hugging Face. Read 19 November 2026. huggingface.co/...Qwen-Image
- ^1 ^2Qwen. "Qwen/Qwen-Image-Edit" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Qwen-Image-Edit
- ^1 ^2lodestones. "lodestones/Chroma1-HD" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Chroma1-HD
- ^1 ^2 ^3 ^4 ^5Tongyi-MAI. "Tongyi-MAI/Z-Image-Turbo" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Z-Image-Turbo
- ^Tongyi-MAI. "Z-Image" (repository README). GitHub. Read 19 November 2026. github.com/...Z-Image
- ^1 ^2Meituan LongCat. "meituan-longcat/LongCat-Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...LongCat-Image
- ^1 ^2 ^3 ^4Z.ai. "zai-org/GLM-Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...GLM-Image
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Black Forest Labs. "FLUX.2 [klein]: Towards Interactive Visual Intelligence". Black Forest Labs, 15 January 2026. bfl.ai/...-towards-interactive-visual-intelligence
- ^1 ^2 ^3 ^4Black Forest Labs. "black-forest-labs/FLUX.2-klein-4B" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.2-klein-4B
- ^1 ^2 ^3 ^4Baidu. "baidu/ERNIE-Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...ERNIE-Image
- ^1 ^2 ^3 ^4HiDream-ai. "HiDream-ai/HiDream-O1-Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...HiDream-O1-Image
- ^1 ^2 ^3NVIDIA. "nvidia/Cosmos3-Super-Text2Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Cosmos3-Super-Text2Image
- ^1 ^2inclusionAI. "inclusionAI/Ming-Image-0.1-Design" (model card and LICENSE). Hugging Face. Read 19 November 2026. huggingface.co/...Ming-Image-0.1-Design
- ^1 ^2OpenMOSS. "MOSS-Image" (repository README). GitHub. Read 19 November 2026. github.com/...MOSS-Image
- ^1 ^2Black Forest Labs. "FLUX.2: Frontier Visual Intelligence". Black Forest Labs, 25 November 2025. bfl.ai/...flux-2
- ^1 ^2 ^3 ^4stable-diffusion-v1-5. "stable-diffusion-v1-5/stable-diffusion-v1-5" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...stable-diffusion-v1-5
- ^1 ^2Stability AI. "Announcing SDXL 1.0". Stability AI, 26 July 2023. stability.ai/...stable-diffusion-sdxl-1-announcement
- ^1 ^2 ^3Stability AI. "CreativeML Open RAIL++-M License dated July 26, 2023" (LICENSE.md, stabilityai/stable-diffusion-xl-base-1.0). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^1 ^2 ^3Stability AI. "Introducing Stable Diffusion 3.5". Stability AI, 22 October 2024 (updated 29 October 2024). stability.ai/...introducing-stable-diffusion-3-5
- ^1 ^2 ^3 ^4Stability AI. "Stability AI Community License Agreement", last updated July 5, 2024 (LICENSE.md, stabilityai/stable-diffusion-3.5-large). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^Tencent. "tencent/HunyuanImage-3.0" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...HunyuanImage-3.0
- ^1 ^2 ^3 ^4Tencent. "Tencent Hunyuan Community License Agreement", Tencent Hunyuan Image 3.0 release date September 28, 2025 (LICENSE, tencent/HunyuanImage-3.0). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3 ^4Tencent. "tencent/HunyuanImage-3.0-Instruct" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...HunyuanImage-3.0-Instruct
- ^1 ^2 ^3 ^4 ^5Krea. "krea/Krea-2-Turbo" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Krea-2-Turbo
- ^1 ^2 ^3 ^4Krea.ai, Inc. "KREA 2 Community License Agreement v.1", 22 June 2026. GitHub (krea-ai/krea-2). Read 19 November 2026. cdn.jsdelivr.net/...LICENSE.pdf
- ^1 ^2 ^3 ^4Black Forest Labs. "FLUX 3 Image Klein: open weights, 5.6B parameters". Black Forest Labs, 15 October 2026. bfl.ai/...flux-3-image-klein
- ^1 ^2 ^3 ^4 ^5 ^6 ^7Black Forest Labs. "black-forest-labs/flux-3-image-klein" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...flux-3-image-klein
- ^1 ^2 ^3 ^4 ^5 ^6Black Forest Labs. "FLUX Kommunity License v.1.0" (LICENSE.md, black-forest-labs/flux-3-image-klein). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^1 ^2 ^3 ^4 ^5 ^6MiniMax. "MiniMaxAI/MiniMax-H3-Image" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...MiniMax-H3-Image
- ^1 ^2 ^3 ^4MiniMax. "MiniMax-H3-Image Community License Agreement", 10 November 2026 (LICENSE, MiniMaxAI/MiniMax-H3-Image). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE
- ^Black Forest Labs. "black-forest-labs/FLUX.1-dev" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.1-dev
- ^1 ^2 ^3Black Forest Labs. "FLUX.1 [dev] Non-Commercial License v1.1.1" (LICENSE.md, black-forest-labs/FLUX.1-dev). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^1 ^2Black Forest Labs. "FLUX.1 Kontext [dev] - Open Weights for Image Editing". Black Forest Labs, 26 June 2025. bfl.ai/...flux-1-kontext-dev
- ^1 ^2 ^3 ^4Black Forest Labs. "black-forest-labs/FLUX.1-Kontext-dev" (model card and LICENSE.md). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.1-Kontext-dev
- ^Black Forest Labs. "FLUX.1 Krea [dev]: An 'Opinionated' Text-to-Image Model". Black Forest Labs, 31 July 2025. bfl.ai/...flux-1-krea-dev
- ^1 ^2Black Forest Labs. "black-forest-labs/FLUX.1-Krea-dev" (model card and LICENSE.md). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.1-Krea-dev
- ^1 ^2Black Forest Labs. "black-forest-labs/FLUX.2-dev" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.2-dev
- ^1 ^2 ^3 ^4 ^5Black Forest Labs. "FLUX Non-Commercial License v2.1" (LICENSE.md, black-forest-labs/FLUX.2-dev). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^1 ^2 ^3 ^4 ^5Black Forest Labs. "black-forest-labs/FLUX.2-klein-9B" (model card and LICENSE.md). Hugging Face. Read 19 November 2026. huggingface.co/...FLUX.2-klein-9B
- ^1 ^2 ^3 ^4Ideogram. "ideogram-ai/ideogram-4-fp8" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...ideogram-4-fp8
- ^1 ^2 ^3 ^4Ideogram. "Ideogram Non-Commercial Model Agreement", last updated June 3, 2026 (LICENSE.md, ideogram-ai/ideogram-4-fp8). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE.md
- ^1 ^2 ^3Qwen. "Qwen/Qwen-Image-2.1" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Qwen-Image-2.1
- ^1 ^2 ^3 ^4QwenLM. "Qwen-Image-2.1" (repository README). GitHub. Read 19 November 2026. github.com/...Qwen-Image-2.1
- ^1 ^2 ^3 ^4Hangzhou Tongyi Laboratory Technology Co., Ltd. "Qwen Research License Agreement", release date September 20, 2026 (LICENSE, Qwen/Qwen-Image-2.1). Hugging Face. Read 19 November 2026. huggingface.co/...LICENSE
- ^1 ^2 ^3 ^4 ^5Bria. "briaai/Fibo-2.0-dev" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...Fibo-2.0-dev
- ^1 ^2Stability AI. "stabilityai/sdxl-turbo" (model card, LICENSE.md and commit history). Hugging Face. Read 19 November 2026. huggingface.co/...main
- ^1 ^2 ^3 ^4 ^5 ^6 ^7GitHub. REST API repository records (licence, stars, last push) for Comfy-Org/ComfyUI, huggingface/diffusers, leejet/stable-diffusion.cpp, invoke-ai/InvokeAI, AUTOMATIC1111/stable-diffusion-webui, sgl-project/sglang, vllm-project/vllm-omni, black-forest-labs/flux-3-image-klein, ideogram-oss/ideogram4 and krea-ai/krea-2. Read 19 November 2026. api.github.com/...ComfyUI
- ^1 ^2Comfy-Org. "ComfyUI" (repository README). GitHub. Read 19 November 2026. github.com/...ComfyUI
- ^1 ^2leejet. "stable-diffusion.cpp" (repository README). GitHub. Read 19 November 2026. github.com/...stable-diffusion.cpp
- ^Stability AI. "License Update". Stability AI, 5 July 2024. stability.ai/...license-update
- ^Runway. "runwayml" (organization page). Hugging Face. Read 19 November 2026. huggingface.co/runwayml
- ^Efficient-Large-Model. "Relicense model under Apache-2.0" (commit d71cbc8, Efficient-Large-Model/Sana_1600M_1024px). Hugging Face, 31 July 2026. huggingface.co/...562b01cd77902952fcf8cbb90aa78e3f
- ^Stability AI. "stabilityai/stable-diffusion-3.5-large" (model card). Hugging Face. Read 19 November 2026. huggingface.co/...stable-diffusion-3.5-large
- ^LAION e.V. "Releasing Re-LAION-5B: transparent iteration on LAION-5B with additional safety fixes". LAION, 30 August 2024. laion.ai/...relaion-5b
- ^Carlini, Nicholas; Hayes, Jamie; Nasr, Milad; Jagielski, Matthew; Sehwag, Vikash; Tramer, Florian; Balle, Borja; Ippolito, Daphne; Wallace, Eric. "Extracting Training Data from Diffusion Models". arXiv:2301.13188, 30 January 2023. arxiv.org/...2301.13188
- ^High Court of Justice (England and Wales), Mrs Justice Joanna Smith. "Getty Images (US) Inc & Ors v Stability AI Limited", [2025] EWHC 2863 (Ch), 4 November 2025. The National Archives. caselaw.nationalarchives.gov.uk/...2863
- ^Getty Images (US), Inc. and others (appellants). "Replacement Skeleton Argument of the Appellants", Court of Appeal ref. CA-2026-000007, 21 October 2026. Courts and Tribunals Judiciary. judiciary.uk/...on--CA-2026-000007-144861071_1.pdf
- ^United States District Court, Northern District of California. "Order Granting in Part and Denying in Part Motions to Dismiss First Amended Complaint", Andersen v. Stability AI Ltd., No. 23-cv-00201-WHO, Document 223, 12 August 2024. CourtListener. storage.courtlistener.com/...cand.407208.223.0.pdf
- ^United States District Court, Northern District of California. "Order Granting in Part Plaintiffs' Motion to Modify Case Schedule", Andersen v. Stability AI Ltd., No. 23-cv-00201-WHO, Document 409, 2 February 2026. CourtListener. storage.courtlistener.com/...cand.407208.409.0.pdf
- ^United States District Court, Northern District of California. "Order Granting Plaintiffs' Motion to Modify Case Schedule Pursuant to Local Rule 6-3", Andersen v. Stability AI Ltd., No. 23-cv-00201-WHO, Document 764, 28 September 2026. CourtListener. storage.courtlistener.com/...cand.407208.764.0.pdf
- ^Hugging Face. "Models Download Stats". Hugging Face Hub documentation. Read 19 November 2026. huggingface.co/...models-download-stats
- ^Black Forest Labs. "FLUX [dev] Non-Commercial License v2.0", last updated November 25, 2025. bfl.ai. Read 19 November 2026. bfl.ai/...non-commercial-license-terms
- ^Black Forest Labs. "black-forest-labs/flux-3-action-base" (model card and LICENSE.md). Hugging Face. Read 19 November 2026. huggingface.co/...flux-3-action-base
Improve this article
Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.
1 revision · v1 · 7,699 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Independent fact-check before publication on November 19, 2026: 318 claims checked against 165 cited and independent sources; 4 material and 13 minor corrections applied before publication.
Cite this page: AI Wiki. "Best open-weight image generation models." aiwiki.ai, updated 19 Nov 2026, fact-checked 19 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/best_open_weight_image_models