.ai

de beste open source AI-modellen

Zoeken

Hyper-SD XL: one-to-eight-step SDXL with a flexible LoRA

Q: Flat, washed-out image at 1 step

Wrong sampler for the 1-step variant. Use TCD with eta 0.3 for 1-2 steps, or move to the 4-step variant with Euler.

Q: Negative prompt is ignored

Standard variant at CFG 1. Use the CFG-preserving variant at CFG 3-5.

Onderwerp: Modellen

Door Captain

Gepubliceerd op 2026-10-02

Hyper-SD from ByteDance accelerates SDXL checkpoints to as few as one step. Learn which variant to pick, the CFG and sampler rules, and how to combine it with checkpoints and LoRAs without losing quality.

Deze pagina is nog niet vertaald en wordt daarom in het Engels getoond.

Overzicht

Hyper-SD is ByteDance's trajectory-segmented consistency distillation. As a

LoRA

it attaches to any

SDXL

checkpoint

and makes it converge in 1, 2, 4 or 8 steps. The catalog's hyper-xl entry is tuned to raise generation speed dramatically while keeping the checkpoint's style.

Compared with SDXL-Lightning, Hyper-SD has a CFG-preserving variant that tolerates a guidance scale of 3-5, which keeps

negative prompts

useful.

Naslag

Naam	Type	Wat het is
LoRA weight	0.8-1.0	1.0 for the plain variant; 0.8 can soften over-sharpening when stacking LoRAs.
Steps	1-8	4 or 8 for quality. The 1-step variant needs the TCD sampler .
CFG	1 or 3-5	1 for the standard LoRA; 3-5 for the CFG-preserving variant.
Sampler	Euler / TCD	Euler with sgm_uniform at 4-8 steps; TCD for 1-2 steps with eta 0.3.
Resolution	1024	Standard SDXL sizes.

Geen installatie nodig – draai AI-modellen in de cloud

BitVector Prism is de makkelijkste manier om te beginnen: kies een model, typ een prompt en genereer in een overzichtelijke web-app, zonder iets in te stellen. BitVector is ook beschikbaar op Discord en op het web (SpyGlass).

BitVector proberen

Stap voor stap

Attach Hyper-SD XL at 1.0 to your SDXL checkpoint.
Choose 8 steps, Euler, sgm_uniform, CFG 1 for your first run.
If you need a negative prompt to work, switch to the CFG-preserving variant and set CFG to 4.
Add your style or character LoRAs at 0.6-0.8.
Use the 1-step variant only for live previews; move to 4-8 steps for anything you keep.

Voorbeelden

Realistic checkpoint, 8 steps

candid photo of a street violinist in a snowy square, warm string lights, shallow depth of field, Fujifilm color

Hyper-SD XL 1.0, 8 steps, CFG 1, Euler, sgm_uniform

CFG-preserving with a negative prompt

illustration of a greenhouse full of carnivorous plants, botanical plate style, soft watercolor

Negative: text, frame, border. Hyper-SD XL CFG variant 1.0, 8 steps, CFG 4

Creative idea: live sketch loop

loose charcoal sketch of a dancer mid-turn, gestural lines, white paper

1-step variant with TCD for live preview, then 8 steps for the keeper

Tips

Hyper-SD keeps more of the checkpoint's color grading than Lightning; it is the better choice for stylized models.
At 8 steps you can run hires fix at 0.35 denoise for a clean 2048 image in seconds.
Two acceleration LoRAs never belong in the same graph.
If detail looks smeared, raise steps before raising CFG.

Probleemoplossing

Flat, washed-out image at 1 step

Waarom het gebeurt

Wrong sampler for the 1-step variant.

Zo los je het op

Use TCD with eta 0.3 for 1-2 steps, or move to the 4-step variant with Euler.

Negative prompt is ignored

Waarom het gebeurt

Standard variant at CFG 1.

Zo los je het op

Use the CFG-preserving variant at CFG 3-5.

Geen installatie nodig – draai AI-modellen in de cloud

PirateDiffusion is alleen op Telegram en gemaakt voor pro's: duizenden modellen, LoRA's en workflows via chatcommando's, met onbeperkt genereren voor een vaste prijs.