.ai

os melhores modelos de IA de código aberto

Buscar

SDXL-Lightning: four-step generation on any SDXL checkpoint

Q: Over-saturated, crunchy images

CFG above 2. Set CFG to 1 and keep Euler with sgm_uniform.

Tema: Modelos

Por Captain

Publicado em 2026-09-30

ByteDance's SDXL-Lightning LoRA turns a 30-step SDXL render into a 4-step one. Learn the step counts, the CFG you must use, the sampler, and how to combine it with your favorite checkpoints and style LoRAs.

Esta página ainda não foi traduzida, por isso é exibida em inglês.

Visão geral

SDXL-Lightning is a progressive adversarial distillation of

SDXL

released by ByteDance. The

LoRA

versions are applied on top of any SDXL

checkpoint

and let it produce a finished image in 2, 4 or 8 steps. The catalog entry is the LoRA, which is the flexible way to use it because it keeps your checkpoint's style.

Because the distillation already decides how the image converges, classic guidance scales break it:

Lightning

wants a

CFG

near 1 and a plain Euler sampler with the sgm_uniform scheduler.

Referência

Nome	Tipo	O que é
LoRA weight	1.0	Use full strength; lowering it reintroduces noise.
Steps	4 or 8	Match the LoRA variant. 4 steps is the sweet spot of speed and quality.
CFG	1-2	1 for the 4-step LoRA; up to 2 for the 8-step variant.
Sampler	Euler	Euler with sgm_uniform. Ancestral or Karras schedulers give blotchy results.
Resolution	1024	Standard SDXL buckets; Lightning does not change the native size.

Sem instalação: rode modelos de IA na nuvem

BitVector Prism é a forma mais fácil de começar: escolha um modelo, digite um prompt e gere em um app web limpo, sem configurar nada. O BitVector também está disponível no Discord e na web (SpyGlass).

Experimentar o BitVector

Passo a passo

Load an SDXL checkpoint you like (realistic, anime or painterly) and attach the Lightning LoRA at 1.0.
Set steps to 4, CFG to 1, sampler Euler, scheduler sgm_uniform.
Prompt as you normally would for that checkpoint;
negative prompts
have little effect at CFG 1, so lean on the positive.
Generate a batch; Lightning is fast enough to explore eight seeds at a time.
For a final image, switch to the 8-step variant at CFG 1.5 for a little more detail.

Exemplos

Fast portrait with a realistic checkpoint

photo of a fisherman mending nets at a harbor, morning fog, warm side light, 35mm, natural skin texture

Steps 4, CFG 1, Euler, sgm_uniform, Lightning LoRA 1.0

Anime checkpoint plus style LoRA

1girl, reading on a train, window light, autumn fields outside, soft colors, detailed background

Lightning LoRA 1.0 plus a style LoRA at 0.7, 8 steps, CFG 1.5

Creative idea: rapid storyboard

storyboard frame, wide shot, a courier cyclist weaving through monsoon traffic, neon reflections on wet asphalt, cinematic anamorphic look

Generate 12 variations and pick the sequence

Dicas

Lightning is the fastest way to audition style LoRAs: stack one at 0.7 and iterate in seconds.
Hands and small faces suffer most at 4 steps; use 8 steps or an inpaint pass for those.
Hires fix still works: 1.5x at 0.4 denoise with the same 4-step settings.
If colors look burnt, your CFG is too high; it is the single most common mistake.

Solução de problemas

Over-saturated, crunchy images

Por que acontece

CFG above 2.

Como corrigir

Set CFG to 1 and keep Euler with sgm_uniform.

Noisy or unfinished output

Por que acontece

Step count lower than the LoRA variant, or LoRA weight below 1.

Como corrigir

Match steps to the variant (4 or 8) and set the weight to 1.0.

Sem instalação: rode modelos de IA na nuvem

O PirateDiffusion é exclusivo do Telegram e feito para profissionais: milhares de modelos, LoRAs e workflows por comandos de chat, com geração ilimitada a preço fixo.