Best Local Image Generation Models: FLUX vs SDXL vs Qwen

For most people in 2026, the best local AI image generation model is FLUX.1 [dev] (12B) for prompt adherence and photorealism, SDXL 1.0 (3.5B) for the deepest LoRA and style ecosystem, and Qwen-Image (20B MMDiT, open-sourced Aug 4 2025) when you need readable text inside the image. If you are short on VRAM or want the fastest generations, Alibaba's Z-Image Turbo (6B, released Nov 26 2025) and FLUX.2 [klein] 4B (Jan 15 2026, Apache 2.0, ~13 GB per its model card) are the speed picks. There is no single winner — the right model depends on whether you care most about prompt accuracy, style variety, text rendering, speed, or VRAM, and this guide ranks each on exactly those axes with verified specs.

Every model here runs fully on your own GPU through ComfyUI, with no cloud, no per-image fees, and no usage logging. The tradeoffs are real though: the newest, most accurate models (FLUX.2 [dev], Qwen-Image) are large, while the lightest models give up some quality or ride on a smaller LoRA library. Let's break it down.

What are the best local AI image models in 2026?

Here is the at-a-glance comparison. Parameter counts, licenses and release dates are taken from each model's official model card or repo; VRAM figures are for the GGUF/fp8 quants most people actually run on consumer GPUs, so treat them as practical minimums, not theoretical floors.

A few things stand out immediately. FLUX.1 [dev] is non-commercial — its weights are free to use but only for non-commercial and non-production work, per Black Forest Labs' license. If you need to sell what you generate, FLUX.1 [schnell], FLUX.2 [klein] 4B, SDXL, SD 3.5 and Qwen-Image are all openly licensed for commercial use, while FLUX.2 [dev] (32B) also carries a non-commercial license. SDXL, despite being the oldest and smallest model here, still has the largest community LoRA and fine-tune library by a wide margin, which is why it refuses to die. The full licence picture is in our guide to local image model licenses for commercial use; if what you need is generation without built-in content filters, the uncensored / unfiltered local models guide covers which open weights people use for that and the terms that still apply.

The newest entry is FLUX 3 Image [klein] (October 15, 2026): one 5.6B checkpoint for both generation and editing, which Black Forest Labs says runs on 8 GB NVIDIA cards with its int8 files in ComfyUI 0.41 or later. Its licence allows free commercial use only below US$5M in annual revenue and with content filtering or output review, so read the terms first. Our FLUX 3 Klein local setup guide covers the file sizes, the ComfyUI setup and the licence in plain words.

How were these models ranked?

There is no universal "best image model" score the way HumanEval exists for code, so this comparison ranks on the dimensions that actually decide which model you should download:

• Prompt adherence — does it follow long, specific prompts, including spatial relationships and counts?
• Photorealism / raw quality — how good do faces, skin, lighting and detail look out of the box?
• Text-in-image — can it render readable words, logos and signage without garbling letters?
• Speed — how many steps and how long per image on a typical consumer GPU?
• VRAM — does it fit a 12 GB or 16 GB card via GGUF/fp8 quantization?
• Ecosystem — how many LoRAs, ControlNets and community checkpoints exist?
• License — can you use the output commercially?

The labels below are a qualitative read of what each model card claims and where community use has settled — they are not scores from a benchmark, and we do not attach numbers to them. The speed column is the one with a hard basis: the step count each model card recommends.

Which model is best for photorealism?

FLUX.1 [dev] and FLUX.2 [dev] lead on raw photoreal quality and prompt accuracy. FLUX.1 [dev] is a 12B rectified-flow transformer that, since its August 2024 release, has been the consensus pick for getting a long, detailed prompt rendered faithfully — it follows spatial instructions and complex scenes better than SDXL out of the box. FLUX.2 [dev], released November 25, 2025, is a much larger 32B model that pushes quality and editing further, but it is heavy: Black Forest Labs recommends an H100-class GPU, and locally you realistically need an RTX 4090 with a quantized build. For most people on a single consumer card, FLUX.1 [dev] is the photoreal sweet spot.

The newest contender is Krea 2, whose open weights arrived on June 22, 2026: a 12.9B diffusion transformer in two checkpoints (Raw for fine-tuning, and an 8-step Turbo), with native ComfyUI support from v0.26. It is gated on Hugging Face behind the Krea 2 Community License, which stops being free above a revenue and seat threshold, so read the terms before using it for client work. Setup, VRAM estimates and the licence details are in our Krea 2 local guide.

The catch is the license. FLUX.1 [dev] and FLUX.2 [dev] are both released under non-commercial licenses, so if you plan to sell the images, you want FLUX.1 [schnell] (Apache 2.0, distilled for speed) or FLUX.2 [klein] 4B (also Apache 2.0) instead — both share the FLUX lineage and prompt-following strengths while being commercially usable. For a step-by-step local setup, see our guide to running FLUX locally.

You can confirm the licenses and architecture on the official FLUX GitHub repo and the FLUX.1 [dev] model card.

Which model is best for text inside images?

Qwen-Image wins text rendering, full stop. Alibaba open-sourced Qwen-Image — a 20B Multimodal Diffusion Transformer (MMDiT) — on August 4, 2025, specifically engineered for native text rendering. It handles multi-line layouts, paragraph-level text, posters and signage in both alphabetic languages (English) and logographic ones (Chinese) far more reliably than any FLUX or SD model, where letters tend to garble in longer strings. If your work is graphic-design-adjacent — posters, ads, infographics, anything with words baked into the pixels — Qwen-Image is the model to reach for.

It is also Apache 2.0 licensed (commercial use allowed) and, thanks to community GGUF quants, runs in roughly 12-13 GB at Q4 — fitting a 16 GB card and even squeezing onto smaller ones at heavier quantization. Two follow-ups matter. Qwen-Image-2512 (December 2025) is an Apache 2.0 update of the same 20B model with better human realism and text rendering, per its model card. Qwen-Image-2.0 launched in February 2026 as an API-only model; its open-weights successor, Qwen-Image-2.1 (September 2026), uses a 7B generation component and ships an 8-step Turbo checkpoint, but it is released under the Qwen Research License, which is non-commercial. For commercial text-in-image work, the Apache 2.0 Qwen-Image or Qwen-Image-2512 weights are the ones to download. SD 3.5 Large and the FLUX.2 family also render text noticeably better than older SD models, so they're reasonable runners-up if you're already in those ecosystems.

For a deeper walkthrough, our complete ComfyUI guide covers loading GGUF diffusion models like Qwen-Image step by step.

Which model is best for anime, styles and LoRAs?

SDXL 1.0 is still the undisputed king of style breadth. It's the oldest and smallest model on this list — a 3.5B-parameter base UNet released in July 2023 under the permissive CreativeML OpenRAIL++-M license — but that early, open, commercially-usable release is exactly why it accumulated the largest library of community LoRAs, fine-tuned checkpoints and ControlNets of any local model. Want a specific anime style, a niche aesthetic, a character LoRA, or a particular artist's look? It almost certainly already exists for SDXL and almost certainly does not for FLUX.2 or Qwen-Image yet.

SDXL's raw prompt adherence trails FLUX, and its native text rendering is weak, but for stylized art driven by LoRAs and ControlNet it remains the most flexible, lowest-VRAM (~6-8 GB), most-supported option in 2026. New base models keep arriving, but none has displaced SDXL's ecosystem. To install it, follow our Stable Diffusion (Forge) setup guide.

Which model is fastest / lowest-VRAM?

Two recent releases changed the calculus for people on modest GPUs:

• Z-Image Turbo (6B) — Alibaba's Tongyi Lab released it on November 26, 2025. It's a step-distilled model that produces an image in just 8 inference steps; the model card claims sub-second latency on an enterprise H800 GPU and says it "fits comfortably within 16G VRAM consumer devices". It's Apache 2.0, so commercial use is fine. The undistilled Z-Image base model followed on January 27, 2026 (28-50 steps, with CFG and negative prompts) for LoRA training. The tradeoff: its LoRA ecosystem is still small next to SDXL's.
• FLUX.2 [klein] 4B — Black Forest Labs' January 15, 2026 open release. At 4B parameters its model card puts it at ~13 GB of VRAM with a 4-step distilled sampler, names the RTX 3090/4070 as target consumer GPUs, and describes end-to-end inference "as low as under a second". The 4B models are Apache 2.0 (commercial-safe); the larger [klein] 9B models use BFL's non-commercial license.

Between them, Z-Image Turbo edges ahead on pure photoreal quality per step, while FLUX.2 [klein] 4B has stronger prompt adherence and the FLUX ecosystem behind it. If you want the fastest path to a good image on a 12-16 GB card, start with one of these two. Our Z-Image Turbo in ComfyUI guide walks through the workflow.

One route you may have seen recommended and should skip: both models were briefly runnable with a single ollama run on macOS, but Ollama removed that experimental image runner in v0.32.6 (August 2026) and it has not returned — the models are still listed in its library, and they fail at run time on current builds and never worked on Windows. Ollama image generation: what actually works now has the full timeline; for these two models, ComfyUI is the working path on every OS.

Which app should you run these models in?

The model is only half the choice — it has to load in a local image generator. Here is which front-end runs which family, checked against each project's release notes as of November 2026:

The short version: if you want FLUX.2, Qwen-Image, Z-Image or Krea 2, use ComfyUI or InvokeAI. If you only run SDXL and want the simplest possible local AI image generator, Fooocus still works. For a node-by-node FLUX setup, see the ComfyUI FLUX workflow guide.

How fast are these on real consumer hardware?

We have not benchmarked these ourselves, so rather than publish invented timings, here is what actually determines them. Diffusion cost scales with step count multiplied by model size — a 4-step 4B model does a small fraction of the work of a 28-step 32B one. Both of those columns come from each model's own card, so the ordering below is reliable even though the wall-clock seconds on your GPU are not something we can tell you:

The pattern is clear: distilled/turbo models (Z-Image, klein, schnell) trade a little fidelity for a roughly 4-8x speedup over the big 20-32B models. If you iterate a lot, generate drafts on a turbo model and do final renders on FLUX.1 [dev] or Qwen-Image.

Honest verdict — which should you actually download?

• You want the best prompt adherence and photorealism (non-commercial use): FLUX.1 [dev]. It's still the best all-round local image model for personal projects in 2026.
• You need to sell the output: FLUX.1 [schnell] or FLUX.2 [klein] 4B (both Apache 2.0), SDXL, SD 3.5, or Qwen-Image — avoid the FLUX [dev] models commercially.
• You want readable text, posters, logos: Qwen-Image (20B). Nothing local renders text better.
• You want anime, specific styles, character LoRAs: SDXL 1.0. The ecosystem is unmatched.
• You're on a 12-16 GB card or want speed: Z-Image Turbo or FLUX.2 [klein] 4B.
• You have a 24 GB+ card and want the absolute ceiling: FLUX.2 [dev], if you accept its non-commercial license and slow speed.

The honest summary: SDXL still wins on LoRA and style breadth, FLUX wins on prompt adherence and photorealism, and Qwen-Image wins on text. No single model dominates all three.

Key Takeaways

1. FLUX.1 [dev] (12B) is the best all-round local image model in 2026 for prompt adherence and photorealism — but its license is non-commercial.
2. For commercial use, pick an openly licensed model: FLUX.1 [schnell] and FLUX.2 [klein] 4B (Apache 2.0), SDXL (OpenRAIL++-M), SD 3.5 (Community License), or Qwen-Image (Apache 2.0).
3. Qwen-Image (20B MMDiT, Aug 2025) is the text-rendering champion — use it for posters, signage and anything with words in the image.
4. SDXL 1.0 (3.5B, Jul 2023) still wins style and LoRA breadth despite being the oldest and smallest model here.
5. Z-Image Turbo (6B, 8 steps) and FLUX.2 [klein] 4B (4 steps) are the speed/low-VRAM picks, both running on 12-16 GB cards.

Next Steps

• New to local image generation? Start with our complete ComfyUI guide — it's the front-end every model here loads into.
• Setting up FLUX specifically? Read how to run FLUX locally for the dev/schnell setup and recommended quants.
• Want the fastest workflow on a modest GPU? Follow our Z-Image Turbo in ComfyUI walkthrough.
• Prefer the classic Stable Diffusion stack with the biggest LoRA library? Use the Stable Diffusion Forge guide to install SDXL.