whatmodel.sh / image / Stable Diffusion XL 1.0

Stable Diffusion XL 1.0

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Stable Diffusion XL 1.0 is an open-weights image model from Stability AI, released July 2023. Its published size is 3.5B parameters. Licence: openrail++.

34.1±1.3#122 of 144 image modelsfrom 15 benchmark results across 14 sources

Score within the LMArena text-to-image comparison set only.

$ how it compares

The image models ranked either side of Stable Diffusion XL 1.0, out of 197 with a score, then more from Stability AI.

#modelscorewhereRAMreleased
165Ovis-U1-3B Alibaba (AIDC-AI)35±6comparison only2025-06
166IF-XL DeepFloyd35±2local~29GB2023-04
167Stable Diffusion XL 1.0 Stability AI34±1api $0.0019/image +local~4.9GB2023-07
168LCM-SDXL latent-consistency34±3local~8.9GB2023-11
169OmniGen (v1) BAAI34±4local~9GB2024-10
more from Stability AI
137Stable Diffusion 3.5 Large Stability AI42±1api $0.065/image +local~5.4GB2024-10
141Stable Diffusion 3.5 Medium Stability AI41±3api $0.035/image +local~20GB2024-10
148Stable Diffusion XL 0.9 (base) Stability AI40±6local~4.9GB2023-06

see all image models · find the ones that fit your machine

$ what it's good at

text to image▂▅▅▃▄▄█▃▁▂▁▁ #106 of 126 · 82% of its score · 12 tests

Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), IF-XL here, though they rank higher overall. Trails SDXL-Turbo here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
GenAI-Arena text-to-image (Elo, 4 Jun 2024)1001.0 ✎▃▃··▃·▃·█·▃▃ beats 3/8avg 1237▇▇▇▇▇ 13%sharp
EvalMuse-40K model ranking (FGA-BLIP2 overall alignment)2.99 / 5 ✎█▄·▄██▄▄█▄▄▄ beats 8/14easy · avg 3.83 / 5▇▇▇▇▇ 12%at its level
GenEval55.0% ✎▅█···▅▃▃▃··▃ beats 9/10easy · avg 79%▇▇▇▇▇ 11%at its level
ImagenWorld text-guided generation (human overall)64.0% ✎▄··█▄·▄··▄·▄ beats 0/6easy · avg 84%▇▇▇▇ 10%at its level
ConceptMix (k=2 full-mark score)43.0% ✎▄·▄·█▄▄▄···▄ beats 5/7easy · avg 66%▇▇▇▇ 10%at its level
ImagenHub text-guided generation (human overall)59.0% ✎▄▄·█·······▄ beats 2/4easy · avg 80%▇▇ 5.4%at its level
GeckoNum Task 1: exact number generation (human-annotated)22.8% ✎██·█·······█ beats 1/3medium · avg 51%▇▇ 5.1%typical
ImageTime (prompt-only, overall)15.7% ✎▄·····▄█▄·▄█ beats 0/7medium · avg 41%▇ 3.7%typical
4 more tests
testresultvs the others in its rundifficultycountswhy
Human text alignment · wins against Kandinsky 3.035.0% ✎█·█········█ beats 1/2medium · avg 58%▇ 3.6%at its level · small run
Human visual quality · wins against Kandinsky 3.041.0% ✎█····█·····█ beats 0/2easy · avg 66%▇ 3.5%at its level · small run
MMMG knowledge-image generation (MMMG-Score, average over education levels)15.9% ✎▃··▃·█▃▃▅··▃ beats 6/9hard · avg 24%▇ 3.2%typical
HPSv3 image generation benchmark (Overall HPSv3 score)8.2000 ✎▃·····▃··▃·█ beats 1/5avg 10.058▇ 0.6%sharp · noisy
other tests3 not in a task · 17% of its score
testresultvs the others in its rundifficultycountswhy
EvalAlign human text-image alignment score5.0300 mean human alignment score ✎█▄▄·█·▄█▄▄·▄ beats 6/11avg 6.285 mean human alignment score▇▇▇▇ 11%typical
MS-COCO 10K FID (whole image) ↓ (SDXL-Lightning protocol)18.4900 FID ✎█··██······█ beats 3/3avg 11.695 FID▇▇ 3.8%typical
VVRBench accuracy (10,000 verifiable synthetic-scene tasks)0.0% ✎▃·█▅▃·▅▃···▃ beats 0/10very hard · avg 7.6%▇ 2.3%sharp · too hard for it

$ where to run it

Through an API

api · Together AI · $0.0019/image · www.together.ai ↗ · price sourceUSD for one 1-megapixel output: Together bills stabilityai/stable-diffusion-xl-base-1.0 at $0.0019 per output megapixel (price = rate × 1 MP; a 1024×1024 image is ~1.05 MP). Together states prices include default steps and extra steps cost more; the default step count for SD XL is not listed. Product: Together AI serverless image generation. API model: stabilityai/stable-diffusion-xl-base-1.0. Settings: 1 MP output, n=1, default steps. Price basis: published per-megapixel list rate × 1 MP. Currency: USD. Replaces a Replicate link whose SDXL endpoint is billed by hardware time only. Together AI official model page, observed 2026-09-25.

Other ways to run it

version

  1. Run it as published · needs ~8.9 GB, NVIDIA only · from stability.ai (official)
    uv run --with diffusers --with transformers --with accelerate --with torch --with sentencepiece python -c "import torch; from diffusers import DiffusionPipeline; p=DiffusionPipeline.from_pretrained('stabilityai/stable-diffusion-xl-base-1.0', torch_dtype=torch.float16, use_safetensors=True, variant='fp16').to('cuda'); p('a cat holding a sign').images[0].save('out.png')"

Official weights and documentation: huggingface.co/stabilityai ↗

$ how the score was worked out · GenAI Arena: An Open Evaluati… 13%, EvalMuse-40K: A Reliable and… 12%, GenEval authors' results: djg… 11% and 11 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

sourceresultsshare of the scorelicence
GenAI Arena: An Open Evaluation Platform for Generative Models (arXiv v3)1▇▇▇▇ 13%CC-BY-4.0
EvalMuse-40K: A Reliable and Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Evaluation (arXiv v2)1▇▇▇▇ 12%CC-BY-4.0
GenEval authors' results: djghosh13/geneval README and arXiv 2310.11513v11▇▇▇ 11%MIT (README); CC-BY-4.0 (arXiv 2310.11513v1)
EvalAlign: Supervised Fine-Tuning Multimodal LLMs with Human-Aligned Data for Evaluating Text-to-Image Models (arXiv:2406.16562v3)1▇▇▇ 11%CC-BY-4.0
ImagenWorld: Stress-Testing Image Generation Models with Explainable Human Evaluation on Open-ended Real-World Tasks1▇▇▇ 10%CC-BY-4.0
ConceptMix: A Compositional Image Generation Benchmark with Controllable Difficulty (arXiv v1)1▇▇▇ 10%CC-BY-4.0
Kandinsky 3.0 — original human comparisons2▇▇ 7%CC BY 4.0
ImagenHub: Standardizing the evaluation of conditional image generation models (arXiv v4)1▇▇ 5%CC-BY-4.0
Evaluating Numerical Reasoning in Text-to-Image Models (GeckoNum, arXiv v3)1▇▇ 5%CC-BY-4.0
SDXL-Lightning: Progressive Adversarial Diffusion Distillation (arXiv:2402.13929v3)1▇ 4%CC-BY-4.0
Can Image Models Imagine Time? ImageTime: A Novel Benchmark for Probing Visual World Modeling Through Spatiotemporal Consistency1▇ 4%CC-BY-4.0
MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (arXiv:2506.10963v2)1▇ 3%CC-BY-SA-4.0
Verifiable Visual Rewards Transfer from Synthetic Scenes to Natural Prompts (VVRBench), arXiv:2609.35641v11▇ 2%CC-BY-4.0
HPSv3 repository README, image generation benchmark table (MizzenAI/HPSv3)1▇ 1%MIT
prior · pull toward the middle▇ 1%
weighted average= 34.1
$ published speed measurements · 4

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

MacBook Pro (M1 Max)fp1646 s/imageCore ML StableDiffusion Swift pipeline · 1024x1024; 20 steps; classifier-free guidance; UNet batch=2; CPU_AND_GPU; ORIGINAL attention; median of 3 end-to-end runs

Apple and Hugging Face July 2023 benchmark on public OS betas. UNet batch=2 implements guidance, not two output images.

source ↗
MacBook Pro (M2 Max)fp1637 s/imageCore ML StableDiffusion Swift pipeline · 1024x1024; 20 steps; classifier-free guidance; UNet batch=2; CPU_AND_GPU; ORIGINAL attention; median of 3 end-to-end runs

Apple and Hugging Face July 2023 benchmark on public OS betas. UNet batch=2 implements guidance, not two output images.

source ↗
Mac Studio (M1 Ultra)fp1625 s/imageCore ML StableDiffusion Swift pipeline · 1024x1024; 20 steps; classifier-free guidance; UNet batch=2; CPU_AND_GPU; ORIGINAL attention; median of 3 end-to-end runs

Apple and Hugging Face July 2023 benchmark on public OS betas. UNet batch=2 implements guidance, not two output images.

source ↗
Mac Studio (M2 Ultra)fp1620 s/imageCore ML StableDiffusion Swift pipeline · 1024x1024; 20 steps; classifier-free guidance; UNet batch=2; CPU_AND_GPU; ORIGINAL attention; median of 3 end-to-end runs

Apple and Hugging Face July 2023 benchmark on public OS betas. UNet batch=2 implements guidance, not two output images.

source ↗

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated · sources and licences · json