whatmodel.sh / image / SDXL-Lightning (4-step UNet)
SDXL-Lightning (4-step UNet)
↗ share
+ contribute
SDXL-Lightning (4-step UNet) is an open-weights image model from ByteDance, released February 2024. Its published size is 2.57B parameters. Licence: openrail++.
Catalogue identity notes
ByteDance/SDXL-Lightning sdxl_lightning_4step_unet.safetensors on SDXL 1.0 base (the 1/2/8-step checkpoints are distinct and not mapped). Memory estimated as for SDXL 1.0.
Model information sources
UNet only: OUR COUNT from the safetensors header of the official 4-step UNet checkpoint (2,567,463,684 F16; SDXL UNet architecture). The HF API gives no total for this multi-checkpoint repo. The two CLIP text encoders and the VAE are not included. Observed 2026-10-07.
Score within the LMArena text-to-image comparison set only.
$ how it compares
The image models ranked either side of SDXL-Lightning (4-step UNet), out of 197 with a score, then more from ByteDance.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 154 |  NextStep-1-Large-Edit | 37±3 | comparison only | ||
| 155 | Step1X-Edit v1.0 | 37±4 | comparison only | ||
| 156 | SDXL-Lightning (4-step UNet) | 37±3 | local | ||
| 157 |  Kandinsky 2.2 | 37±5 | local | ||
| 158 |  Janus-Pro-1B | 36±4 | api $0.0005/image +local | ||
| more from ByteDance | |||||
| 30 | Seedream 4.5 | 64±1 | api $0.040/image | ||
| 129 | UNO | 43±4 | local | ||
$ what it's good at
text to image▂▅▅▃▄▄█▃▁▂▁▁ #93 of 126 * · 74% of its score · 1 test
Beats Latent Diffusion text2img-large (LDM-KL-8-G, LAION-400M), VQ-Diffusion (ITHQ), Stable Diffusion XL 0.9 (base) here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GenAI-Arena text-to-image (Elo, 4 Jun 2024) | 1053.0 ✎ | ▃▃··▃·▃·█·▃▃ beats 5/8 | 74% |
other tests1 not in a task · 22% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| MS-COCO 10K FID (whole image) ↓ (SDXL-Lightning protocol) lab | 22.3000 FID ✎ | █··██······█ beats 1/3 | 22% |
$ where to run it
Other ways to run it
version
- Run it as published · needs ~8.9 GB, NVIDIA only · from huggingface.co/ByteDance (official)
uv run --with diffusers --with transformers --with accelerate --with torch --with safetensors python -c "import torch; from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler; from huggingface_hub import hf_hub_download; from safetensors.torch import load_file; b='stabilityai/stable-diffusion-xl-base-1.0'; u=UNet2DConditionModel.from_config(b, subfolder='unet').to('cuda', torch.float16); u.load_state_dict(load_file(hf_hub_download('ByteDance/SDXL-Lightning', 'sdxl_lightning_4step_unet.safetensors'), device='cuda')); p=StableDiffusionXLPipeline.from_pretrained(b, unet=u, torch_dtype=torch.float16, variant='fp16').to('cuda'); p.scheduler=EulerDiscreteScheduler.from_config(p.scheduler.config, timestep_spacing='trailing'); p('a cat holding a sign', num_inference_steps=4, guidance_scale=0).images[0].save('out.png')"
Official weights and documentation: huggingface.co/ByteDance ↗
$ how the score was worked out · GenAI Arena: An Open Evaluati… 74%, SDXL-Lightning: Progressive A… 22%
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
| source | results | share of the score |
|---|---|---|
 GenAI Arena: An Open Evaluation Platform for Generative Models (arXiv v3) | 1 | 74% |
| SDXL-Lightning: Progressive Adversarial Diffusion Distillation (arXiv:2402.13929v3) · lab-reported | 1 | 22% |
| prior · pull toward the middle | 4% | |
| weighted average | = 36.9 |
$ improve this page
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