Best open-source image generation models (November 2026)

How this list is put together

There's no single best open-source image model. The right one depends on what you're making, what GPU you have and whether you can ship the output commercially. So this list sorts models by use, and every fact in it comes from the developer's own model card, repo or launch post, linked at the end.

Four things decide where a model lands:

• Licence. Apache-2.0 and MIT allow commercial use. Several 2026 releases use research or non-commercial licences, which rules them out for most products.
• Size and VRAM. Parameter count and the VRAM figure the developer publishes. Many cards give no VRAM figure at all, and the table says so rather than guessing.
• Capabilities. Text to image only, editing only, or both in one model; multi-reference input; text rendering as the developer describes it.
• Release date and tooling. Newer isn't automatically better, but it matters for whether LoRA trainers and ComfyUI support the model yet.

Not every model here is open source. We call a model open-source only when its weights are under an OSI-approved licence (Apache-2.0 or MIT). The rest are open-weight: you can download the weights, but the licence restricts use, usually commercial use. They're included because they're often what people mean by open-source image models, and the table marks which is which. Our commercial use page has the full definition.

Summary table

Parameter counts are for the image model unless noted. Most of these also load a separate text encoder, which adds to the memory you need: FLUX.2 [klein] 9B pairs with an 8B Qwen3 text embedder, and FLUX.3 Image klein uses Qwen3-VL-4B-Instruct.

Best for general text to image

Z-Image Turbo

A 6B single-stream diffusion transformer from Alibaba's Tongyi-MAI lab, distilled to 8 function evaluations. The card says it fits within 16 GB consumer GPUs and lists photorealism, bilingual (English and Chinese) text rendering and instruction adherence as its strengths. It's Apache-2.0 with no extra terms. The trade-off is in its own family's comparison table: no CFG, no negative prompts, low diversity across seeds, and not meant for fine-tuning. Settings from the card: 9 inference steps (8 DiT forward passes) and guidance_scale=0.0.

Qwen-Image-2512

The December 2025 update of Qwen-Image, a 20B MMDiT model the Qwen team built around complex text rendering, especially Chinese. The card lists three changes from the August release: more realistic people, finer natural detail such as fur and landscapes, and more accurate text and layout. It's Apache-2.0. It's much larger than Z-Image Turbo and runs 50 steps with true_cfg_scale=4.0 in the card's example, with a negative prompt. The Qwen README says it ranked as the strongest open-source image model in over 10,000 blind rounds on Alibaba's AI Arena; that's the developer's claim.

HiDream-O1-Image

An 8B model from HiDream.ai, released in May 2026, that works on raw pixels with no separate VAE or text encoder. It generates directly at up to 2048 × 2048 and also edits and personalises from reference images. It's MIT-licensed. The card asks for a CUDA GPU and flash-attn but gives no VRAM figure.

MOSS-Image

Released in October 2026 by OpenMOSS as two checkpoints, Base and SFT, both Apache-2.0 for weights and code. Besides plain prompts it accepts a JSON layout as the prompt, with bounding boxes for text and objects. Defaults: 50 steps, CFG 3, shift 5. The README reports 3.65 GiB peak GPU memory and 14 GiB system RAM for one 1024 × 1024 image with its memory-saving modes, which keep the DiT in CPU RAM; it doesn't give a figure for keeping the whole model on the GPU.

Z-Image, the undistilled base, belongs here too if you want negative prompts and more variety: 28 to 50 steps, guidance 3 to 5, and the same Apache-2.0 licence.

Best open-source image edit model

Editing models take one or more images plus an instruction. The commercially usable options are narrower than for text to image.

Qwen-Image-Edit-2511 is the default pick for instruction editing under a permissive licence. Its card lists less image drift, better character and multi-person consistency, and some popular community LoRA effects (lighting, new viewpoints) built in. The card's example runs 40 steps with true_cfg_scale=4.0.

FLUX.2 [klein] 4B is the option when you want one small model for both generation and editing. It's Apache-2.0, and the card puts it at about 13 GB.

Qwen-Image-2.1 has the broadest editing feature list on paper: up to 10 references, local edits marked with circles, painted annotations or masks, and native RGBA (transparent) output. But it's open-weight, under the Qwen Research License, which allows non-commercial use only, so treat it as something to evaluate rather than ship.

Best for low VRAM

If you're searching for open-source or open-weight image models with VRAM requirements, the short answer is that few developers publish one. These are the ones that do, as they state them:

For a 12 to 16 GB card under a commercial licence, FLUX.2 [klein] 4B and Z-Image Turbo are the two with published figures that fit. BFL also ships official FP8 and NVFP4 versions of every [klein] model and says they use up to 40% and 55% less VRAM respectively.

The 20B Qwen models have no official VRAM figure. The Qwen README points to DiffSynth-Studio, a community tool, for layer-by-layer offload that it says runs within 4 GB, and FP8 quantisation. Our hardware requirements page collects each catalogued model's published figures in one place.

Best for commercial use

These models carry licences that allow commercial use without a separate agreement:

• Apache-2.0 (open-source): Z-Image Turbo, Z-Image, Qwen-Image-2512, Qwen-Image-Edit-2511, FLUX.2 [klein] 4B and 4B Base, FLUX.1 [schnell], MOSS-Image, LongCat-Image.
• MIT (open-source): HiDream-O1-Image, GLM-Image (GLM-Image bundles tokenizer and ViT weights that stay under Apache-2.0).
• Open RAIL++-M (open-weight): SDXL. No non-commercial clause, but its use restrictions apply, and if you host it as a service you have to put them into your terms with users.

Watch the family names. Within FLUX.2, only the two 4B [klein] models are Apache-2.0; the 9B models and FLUX.2 [dev] are under BFL's non-commercial licence. Within FLUX.1, only [schnell] is Apache-2.0. Within Qwen-Image, the 2025 releases are Apache-2.0 and Qwen-Image-2.1 is not.

FLUX.3 Image klein, also open-weight, sits in between. Its FLUX Kommunity License v1.0 allows non-commercial use plus commercial output generation by qualifying users with less than US$5M in gross annualised revenue, affiliates included. Read the licence itself before relying on that. Our commercial use page covers the conditions for each catalogued model.

Best for LoRA training

Distilled models (Turbo, [klein] without "Base", [schnell]) are tuned for few-step inference and aren't the intended base for training. Most families now publish an undistilled version for that:

• Z-Image (base). Its card calls it the starting point for LoRA training and ControlNet-style conditioning.
• FLUX.2 [klein] Base 4B (Apache-2.0) and Base 9B (non-commercial). BFL's README recommends the Base models, 50 steps, for fine-tuning and LoRA training.
• Qwen-Image / Qwen-Image-2512, trained directly.
• MOSS-Image-Base, with LoRA training guidance in its user guide.
• SDXL, the oldest model here, with a large library of community LoRAs, fine-tunes and ControlNets built up since 2023.

On tooling: ai-toolkit lists Z-Image Turbo, Z-Image, FLUX.2 [klein] Base 4B and 9B, FLUX.3 Image klein Base, Qwen-Image, Qwen-Image-2512, the Qwen-Image-Edit releases, Qwen-Image-2.1, HiDream-O1, MOSS-Image, Ideogram 4 and SDXL among its supported models. musubi-tuner supports FLUX.2 dev and [klein], Qwen-Image, Z-Image and HiDream-O1-Image. For Z-Image, musubi-tuner's docs recommend training on the base model now that it's out.

What's new in 2026

• January: FLUX.2 [klein] 4B and 9B, with Base versions and FP8/NVFP4 builds. Z-Image base. GLM-Image, a 9B autoregressive generator plus a 7B diffusion decoder.
• March: FLUX.2 [klein] 9B KV, which BFL says is faster than 4B for multi-reference editing through KV caching.
• May: HiDream-O1-Image, followed by its distilled Dev-2604.
• June: Ideogram 4, 9.3B, with weights under a non-commercial licence.
• September: Qwen-Image-2.1, one model for generation, editing and transparent images, under a research licence.
• October: Qwen-Image-2.1-Turbo (8 steps), FLUX.3 Image klein (5.6B, with a separate 9B prompt upsampler the card says to always run), and MOSS-Image.

The pattern: the 2026 releases add editing and multi-reference input to the same model, and the newest ones move away from Apache-2.0 to open-weight licences. If you need an open-source licence, the strongest options are still mostly from late 2025 and January 2026.

Is FLUX 3 open source?

Partly. Black Forest Labs announced FLUX 3 in July 2026 without weights. The first open FLUX 3 image model is FLUX.3 Image klein, released in October 2026: 5.6B, with Qwen3-VL-4B as its text encoder, up to 10 references and bounding-box layout guidance, under the FLUX Kommunity License (commercial use for qualifying users under US$5M annual revenue). Its card gives peak memory of about 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4 at 1024 × 1024.

We haven't found a statement from BFL on whether the larger FLUX 3 image models will get open weights, so treat anything beyond klein as closed for now.

Leaderboards and rankings

Leaderboards for open-source and open-weight image models are mostly blind human-preference arenas, where people pick the better of two images for the same prompt. Developers cite them in launch posts: Qwen does for Qwen-Image-2512, and Ideogram's card says Ideogram 4 has the best text rendering of any open-weight release it benchmarked. Those are the developers' own claims and measurements.

Arena rankings also move with every release and say nothing about licence, VRAM or whether a model edits. Shortlist on those constraints first, then run your own prompts through the two or three that fit.

Running them through an API

If you'd rather not manage GPUs, Windpaint serves open-source models through one API. Z-Image Turbo and FLUX.2 [klein] 4B are live today for text to image at 1k in ten aspect ratios; Qwen-Image-2512, Qwen-Image-Edit-2511, FLUX.1 [schnell], HiDream-O1-Image and SDXL are listed as coming soon. See the model catalogue, pricing and the image generation docs.