Best Stable Diffusion and Open Image Models

Z-Image Turbo is the model we would install first. Alibaba's Tongyi-MAI lab released it in November 2025 as a distilled (sped-up) version of its 6B-parameter Z-Image model. It needs only 8 sampling steps, and the team says it fits within 16 GB of VRAM on consumer cards and excels at photorealism and English and Chinese text.

The licence is the big advantage. It is Apache 2.0, so you can use it for client work, products and fine-tunes with no revenue cap and no territory limits. The community has adopted it fast: ComfyUI has an official tutorial, ComfyUI's repackaged files show about 7.9 million Hugging Face downloads in the past month, and Alibaba's PAI team has published a ControlNet Union model for structural control. The undistilled Z-Image base model, released in January 2026, is the better starting point for training LoRAs.

Quality is good rather than top tier: it sits at #75 on the Artificial Analysis arena, level with FLUX.2 [klein] 9B and behind Qwen-Image-2512. The model card lists a Z-Image-Edit variant, but we found no public weights for it.

Pick it if you want a fast, legal-to-sell default on a mid-range GPU. Skip it if you need the very best quality and have 24 GB or more.

Qwen-Image-2512 is the highest-quality open model you can use commercially without paying anyone. It is the December 2025 update of Alibaba's 20B-parameter Qwen-Image, and it sits at #42 on the Artificial Analysis arena with an Elo of 998, just 2 points behind FLUX.2 [dev] and ahead of every other Apache or MIT model. Alibaba says the update makes people look less "AI-generated", adds finer natural detail and improves text rendering.

The cost is hardware. ComfyUI's documentation lists the fp8 file at 20.4 GB and tests it on a 24 GB RTX 4090D, where one image takes about 70 to 95 seconds; an 8-step Lightning LoRA roughly halves that. On 16 GB cards you will need smaller quantised versions.

Its sibling Qwen-Image-Edit-2511, also Apache 2.0, handles instruction-based editing, so you get a full open generate-and-edit pair.

Pick it if you have a 24 GB GPU and need images you can sell. Skip it if your card has 12 GB or less; start with Z-Image Turbo or FLUX.2 [klein] 4B.

FLUX.2 [klein] is Black Forest Labs' small, fast FLUX.2, released in January 2026. One model does text-to-image, image editing and multi-reference work (combining several input images), and BFL says it can generate or edit in under half a second on modern hardware. The 4B version runs on cards like an RTX 3090 or 4070 with about 13 GB of VRAM, and BFL has since shipped it on-device on ASUS ProArt laptops.

Read the licence carefully, because the two sizes differ. The 4B model is Apache 2.0, free for commercial use. The 9B model uses the FLUX Non-Commercial License, so paid work needs a licence from BFL. The 9B is clearly better: it ties Z-Image Turbo at #76 on the Artificial Analysis arena, while the 4B sits at #117.

ComfyUI supports both, and BFL offers FP8 and NVFP4 versions that it says run up to 2.7 times faster.

Pick it if you want quick edits and generation on a 12 to 16 GB card, or an Apache 2.0 model you can build into a product. Skip it if raw image quality matters most; Qwen-Image-2512 and FLUX.2 [dev] are clearly better.

HiDream-O1-Image, released in May 2026, takes an unusual approach. It is a single 8B-parameter transformer that works directly on pixels, with no separate VAE (image compressor) or text encoder. The same model does text-to-image, long multilingual text rendering, instruction editing and subject-driven personalisation (keeping a person or product consistent across new scenes) at up to 2,048 x 2,048. Recent updates added layout and skeleton (pose) conditioning.

The MIT licence is the most permissive in this ranking: no revenue cap, no territory limits and no attribution beyond keeping the licence notice.

Quality is strong for its size. The full model sits at #60 on the Artificial Analysis arena (Elo 979), ahead of Z-Image Turbo. HiDream said a Dev version debuted at #8 in May; leaderboards move fast.

The weakness is tooling. It ships with its own inference scripts built on the Transformers library rather than native ComfyUI nodes, and its community is small (about 6,400 Hugging Face downloads in the past month versus hundreds of thousands for Z-Image).

Pick it if you are a developer who wants one permissive model for generation and editing. Skip it if you want a click-and-go ComfyUI workflow.

SDXL is three years old and its base model now ranks near the bottom of the Artificial Analysis arena (#156, Elo 677). So why is it fifth? Because almost nobody uses the base model any more. SDXL is the foundation for a huge library of community fine-tunes, LoRAs and control add-ons, and it runs on 8 GB graphics cards that newer models cannot use.

Every major local app supports it: ComfyUI, Forge, InvokeAI and SwarmUI. Its licence, CreativeML OpenRAIL++-M, allows commercial use as long as you avoid the listed harmful uses. And it is still heavily used: the base repository had about 3.6 million Hugging Face downloads in the past month.

The limits are real. Out of the box it follows complex prompts poorly, struggles with text in images and needs more tries than 2026 models. Fine-tunes close much of that gap for specific styles, not for general prompt following.

Pick it if you have an 8 to 12 GB card, or you rely on a specific SDXL fine-tune or LoRA library. Skip it if you are starting fresh with a 16 GB card; Z-Image Turbo is better in almost every way.

FLUX.2 [dev] is the most capable open model for editing, and the second-best open model for raw quality. Released in November 2025, this 32B-parameter model does text-to-image and editing with up to 10 reference images in one checkpoint, at up to 4 megapixels. It ranks #40 on the Artificial Analysis arena (Elo 1000), and FLUX has one of the largest LoRA and tool ecosystems, with ComfyUI support from launch day.

Two things hold it back. First, hardware: NVIDIA says the full model needs about 90 GB of VRAM to load, and its FP8 version cuts that by 40%; on a 24 GB card you rely on ComfyUI offloading parts to system RAM, which is slower. Second, the licence. The FLUX [dev] Non-Commercial License covers personal, research and testing use only; using it for paid work or in a product needs a BFL licence. BFL claims no ownership of the images, but making them for a client counts as commercial use.

BFL unveiled FLUX 3 in July 2026 but had not released open FLUX 3 image weights by 25 September.

Pick it if you have a large GPU and non-commercial projects. Skip it if you need to sell the results without a licence.

Ideogram 4 is the best open-weight image model on quality. Ideogram released it in June 2026 as its first open model: a 9.3B-parameter diffusion transformer trained from scratch, using the Qwen3-VL-8B vision-language model to understand prompts. On the Artificial Analysis arena its Quality setting ranks #35 with an Elo of 1010, the highest of any open-weights model, and Ideogram's heritage shows in text rendering: posters, logos and signs come out readable.

It also brings design controls that other open models lack: structured JSON prompts, bounding boxes to place elements, colour palettes and native 2K output with aspect ratios up to 6:1.

The licence is the catch. The free weights are for research, evaluation and personal projects only. Commercial use needs the self-serve licence at $300/month for 10,000 images, or Ideogram's own app and API. The weights are gated, and the reference script by default calls Ideogram's hosted prompt-expansion service, so a fully offline setup takes extra work. Community ComfyUI support is young.

Pick it if you want the best open quality and text, for personal use. Skip it if you need free commercial use; choose Qwen-Image-2512.

Stable Diffusion 3.5 is Stability AI's newest open model family, released in October 2024: Large (8.1B parameters), Large Turbo (a 4-step distilled version) and Medium (2.5B, which Stability says needs 9.9 GB of VRAM excluding text encoders). Stability designed it to be easy to fine-tune, and it runs in ComfyUI and Diffusers.

It has aged quickly. SD 3.5 Large sits at #125 on the Artificial Analysis arena (Elo 839), roughly level with FLUX.1 [dev] and well behind the 2025 and 2026 models above it. Its community never grew as large as SDXL's or FLUX's (about 99,000 Hugging Face downloads in the past month for Large, against 3.6 million for SDXL), so there are fewer LoRAs and fine-tunes.

The licence sits in the middle. The Stability AI Community License is free for commercial use only while you and your affiliates earn under $1 million a year; above that you must get an enterprise licence from Stability.

Pick it if you already have SD 3.5 fine-tunes or pipelines. Skip it if you are choosing a model today: Z-Image Turbo and Qwen-Image-2512 are better and use Apache 2.0.

NVIDIA's Cosmos3-Super-Text2Image, released on 31 May 2026, is part of its Cosmos 3 family of "world models" for robotics and physical AI, but it is also a very good general image generator. With a prompt-rewriting (agentic) mode it ranks #46 on the Artificial Analysis arena with an Elo of 995, close to FLUX.2 [dev] and Qwen-Image-2512.

Its licence is excellent. OpenMDW 1.1 allows commercial use, places no restrictions on outputs and has no revenue cap or territory limit. That makes it one of the few near-frontier open models a large company can deploy without a separate deal.

But it is not for home users. At 64B parameters, NVIDIA's recommended serving setup is an 8x H100 server node, and the diffusers example was tested on GB200. There is no mainstream ComfyUI workflow, and the tooling is aimed at physical-AI developers.

Pick it if you run a data centre or rent large GPU servers and need a permissive, high-quality model. Skip it if you have a single consumer GPU; choose Qwen-Image-2512.

HunyuanImage 3.0 is Tencent's largest image model and, Tencent says, the largest open image-generation mixture-of-experts model: 80B parameters in 64 experts, with 13B active for each token (a mixture-of-experts model only uses part of itself for each step). The Instruct version, released in January 2026, adds prompt reasoning and image-to-image editing, and ranks #66 on the Artificial Analysis arena (Elo 963).

Two problems push it to the bottom. The hardware is data-centre class: Tencent's demo is configured for four GPUs by default. And the licence is the most restrictive here. The Tencent Hunyuan Community License does not apply in the European Union, the United Kingdom or South Korea, so you cannot use the model or its outputs there, and services with over 100 million monthly users need a separate licence.

The older HunyuanImage 2.1 (#101, Elo 886) is lighter but carries the same territory limits.

Pick it if you are a research team outside the excluded regions with large GPUs. Skip it if you are in Europe or the UK, or want a model for one PC.