Skip to content

SDXL

Fast and flexible. Dual-CLIP text encoding (CLIP-L + CLIP-G) with UNet denoising and classifier-free guidance.

Variants

ModelStepsSizeNotes
sdxl-turbo:fp1645.1 GBUltra-fast, 1–4 steps
dreamshaper-xl:fp1685.1 GBFantasy, concept art
juggernaut-xl:fp16305.1 GBPhotorealism, cinematic
realvis-xl:fp16255.1 GBPhotorealism, versatile
playground-v2.5:fp16255.1 GBArtistic, aesthetic
sdxl-base:fp16255.1 GBOfficial base model
pony-v6:fp16255.1 GBAnime, art, stylized
cyberrealistic-pony:fp16255.1 GBPhotorealistic Pony

Defaults

  • Resolution: 1024x1024
  • Guidance: 7.5 (0.0 for turbo)
  • Scheduler: DDIM (also supports euler-ancestral, uni-pc)
WidthHeightAspect Ratio
102410241:1 (native)
11528969:7
89611527:9
121683219:13
832121613:19
13447687:4
76813444:7
153664012:5
64015365:12

Using non-recommended dimensions will trigger a warning. All values must be multiples of 16.

Example

SDXL Turbo — 4 steps, seed 88:

bash
mold run sdxl-turbo:fp16 "A vibrant street food market in Bangkok at night, neon signs, steam from woks, bustling crowd" --seed 88

Street market — SDXL Turbo

Negative Prompts

SDXL uses classifier-free guidance — negative prompts have a strong effect:

bash
mold run sdxl-base:fp16 "a landscape" -n "low quality, blurry, watermark"