SDXL-Lightning
Progressive adversarial distillation of SDXL that renders 1024px images in as few as 2-4 steps.
About
SDXL-Lightning is ByteDance's distilled version of Stable Diffusion XL, built with progressive adversarial diffusion distillation (paper arXiv 2402.13929, February 2024). It ships checkpoints tuned for 1, 2, 4, and 8 inference steps, with the 2-step and 4-step versions producing what the authors call amazing quality at 1024px in under a second on an A100. ByteDance released both full UNet and LoRA checkpoints, and the LoRA variants drop onto any SDXL-based model, which is why the anime and stylized communities adopted it as a near-universal speed plugin. It also slots into ComfyUI and works with ControlNet.
The model came out under an openrail++ license as part of ByteDance's research releases, with demos including the fastsdxl.ai real-time generator. It sits in a crowded speed race against SDXL Turbo and LCM-LoRA, and the fair criticism is that quality at 1 step is experimental and that no standardized benchmark scores were published, the paper's case is qualitative side-by-sides against SDXL and SDXL Turbo. Practically, the 4-step UNet and the LoRA plugin remain the default answer when someone asks how to make SDXL fast.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
No published standard benchmark scores; paper compares outputs qualitatively against SDXL and SDXL Turbo at 1-8 steps
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: ByteDance/SDXL-Lightning Hugging Face card (openrail++ license, 1/2/4/8-step checkpoints, diffusers library); paper arXiv 2402.13929; pandaily and openlaboratory coverage of the release. Strengths: genuinely fast open-weights distillation, LoRA plugin works on any SDXL base, wide community adoption including ComfyUI and ControlNet. Docked for: (1) no standardized benchmark scores were published, quality claims rest on the paper's qualitative comparisons; (2) ByteDance's research-release posture means no maintained product roadmap. License: openrail++.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ByteDance; released 2024-02; paper arXiv 2402.13929
- · 1/2/4/8-step distilled checkpoints (full UNet + LoRA)
- · No published standard benchmark scores — quality claims rest on the paper's qualitative comparisons
- · Third-party APIs (e.g. Runflow) from ~$0.003/image