SDXL Lightning and SDXL Turbo are both distilled variants of Stable Diffusion XL that dramatically cut step count. On our dedicated GPU hosting they make SDXL throughput competitive with models an order of magnitude smaller. Which one to pick depends on your quality-speed trade.
Contents
The Two
- SDXL Turbo: Stability AI’s 1-step distillation. Fastest option. Quality at 1 step is good, 2-4 steps marginal gains.
- SDXL Lightning: ByteDance’s 1/2/4/8-step distillations. More quality at higher step counts, more flexibility.
Speed
1024×1024 on a 5090:
| Variant | Steps | Time/image |
|---|---|---|
| SDXL base | 30 | ~1.4s |
| SDXL Lightning 8-step | 8 | ~0.4s |
| SDXL Lightning 4-step | 4 | ~0.22s |
| SDXL Lightning 2-step | 2 | ~0.13s |
| SDXL Turbo | 1 | ~0.08s |
Quality
Subjective rating on photorealistic prompts (1-10):
| Variant | Quality |
|---|---|
| SDXL base 30 steps | 9 |
| Lightning 8-step | 8.5 |
| Lightning 4-step | 8 |
| Lightning 2-step | 7 |
| Turbo 1-step | 6.5 |
Lightning at 8 steps is nearly indistinguishable from base SDXL while running 3-4x faster. Turbo is noticeably rougher but produces images in under 100ms.
Pick
For production image generation where quality matters: Lightning 8-step. For real-time interactive demos or massive throughput: Turbo. Base SDXL at 30 steps is mostly obsolete once Lightning 8 gets you 95% of the quality at 3x speed.
