.ai

les meilleurs modèles d'IA open source

Rechercher

Steps, CFG, samplers, schedulers and seeds: the generation settings explained with safe defaults

Q: Blurry or mushy result

Too few steps for a non-distilled model, or CFG too low. 25 to 30 steps, CFG at the family default.

Q: Image changes completely when I change steps

An ancestral or SDE sampler. Switch to Euler or DPM++ 2M if you want stable refinements.

Sujet : Prompting

Par Quartermaster

Publié le 2025-11-19

What each setting in the generation panel changes, the values that work per model family (SD 1.5, SDXL, Flux, Qwen, Z-Image, Wan), which samplers are worth knowing, why distilled models break the usual rules, and how to use seeds to change one thing at a time.

Cette page n'est pas encore traduite ; elle est affichée en anglais.

Aperçu

The generation panel has five numbers that everyone fiddles with and few understand. Steps is how many denoising iterations run. CFG (or guidance) is how strongly the prompt steers each iteration. The sampler is the numerical method that turns the model's noise prediction into the next image state; the scheduler is the curve of how much noise is removed per step. The seed is the random starting noise, and fixing it is what lets you compare any other change fairly.

The right values depend on the model family far more than on the picture.

SD 1.5

and

SDXL

are classic diffusion models that like 20 to 30 steps and a CFG of 5 to 7.

Flux dev

is guidance-distilled: CFG stays at 1 and a separate guidance value of 3 to 4 does the steering. Turbo, Lightning, Schnell, Klein and

Z-Image Turbo

are step-distilled: 4 to 8 steps, CFG 1, and more steps make them worse. Video models (

Wan 2.2

LTX-2

) add frame counts and often a two-stage high-noise/low-noise split. Every model page on this site shows the settings under its example images; that is the fastest way to get the numbers for a specific model.

The same numbers apply on the cloud:

PirateDiffusion

exposes them as /steps, /guidance, /sampler and /seed flags;

BitVector

shows them in an advanced panel with the model's defaults preloaded.

Référence

Nom	Type	Rôle
Steps	SD 1.5: 20-30 \| SDXL: 25-35 \| Flux dev: 20-28 \| distilled: 4-8	More steps past convergence add time, not quality. Distilled models get worse past their trained count.
CFG scale	SD 1.5: 6-8 \| SDXL: 5-7 \| Flux dev: 1 \| Qwen: 3-4 \| distilled: 1-2	Too high: burnt colours, oversharpened edges, repeated elements. Too low: muddy, ignores the prompt.
Guidance (Flux)	3-4 (dev), 2.5 for realism LoRAs	The distilled replacement for CFG. Above 5 the plastic Flux look appears.
Sampler	Euler, Euler a, DPM++ 2M, DPM++ SDE, UniPC, DEIS, LCM	Euler and DPM++ 2M are deterministic and converge cleanly; "a" (ancestral) and SDE samplers add noise each step and never fully converge, which gives variety but changes with step count.
Scheduler	normal/simple, Karras, exponential, sgm_uniform, beta	The noise curve. Karras for SD 1.5 and SDXL; simple or beta for Flux; sgm_uniform for Lightning and some video models.
Seed	integer	Same seed + same settings = same image on the same machine. -1 or random picks a new one per run.
Denoise	0-1 (img2img, hires-fix)	How much of the input image is replaced. 1.0 ignores it; 0.3 keeps it almost intact.
Shift (Flux, Wan, Z-Image)	1-8 (model dependent)	Moves where the schedule spends its steps; higher shift favours composition over detail. Model pages list it when it matters.

Defaults preloaded per model

BitVector Prism loads each model with its tested steps, guidance and sampler, so the first image already looks like the examples. Open the advanced panel when you want to experiment.

Essayer BitVector

Pas à pas

Copy the settings line from an example image on the model page; it is the author's tested combination.
Fix the seed to any number while you work on the prompt, so that each change you see is caused by the prompt and not by the noise.
Set steps to the family default and leave it; do not chase quality by adding steps.
Set CFG or guidance to the family default. If the image looks burnt or oversaturated, lower it by one; if it ignores the prompt, raise it by one.
Choose Euler or DPM++ 2M with the family's scheduler. Try an ancestral or SDE sampler only when you want more variation between seeds.
Once the prompt is right, switch the seed to random and generate a batch of four to eight; pick the best composition.
Lock that seed again for refinements (hires-fix,
inpainting
,
LoRA
weight changes) so the picture stays recognisable while you improve details.

Exemples

Reference settings per family

SD 1.5: 512x768 | DPM++ 2M Karras | 25 steps | CFG 7

SDXL: 1024x1024 | DPM++ 2M Karras | 30 steps | CFG 6

Flux dev: 1024x1024 | Euler simple | 24 steps | CFG 1 | guidance 3.5

Z-Image Turbo: 1024x1024 | Euler | 8 steps | CFG 1

PirateDiffusion flags

/render <flux-dev> a lighthouse at dusk, long exposure /steps:24 /guidance:3.5 /seed:123456 /size:1024x1024

Astuces

Steps and CFG interact: at higher CFG you need slightly more steps to avoid artefacts; at lower CFG fewer steps look fine.
Ancestral samplers (Euler a, DPM++ 2S a) change the image completely with each step count change; deterministic ones (Euler, DPM++ 2M) only refine it.
For Flux dev at 20 steps with guidance 3.5, Euler + simple or beta is the reference combination; Karras hurts Flux.
SDXL Lightning and Hyper-SD LoRAs let you run a normal SDXL
checkpoint
at 4 to 8 steps with CFG 1 to 2; the LoRA guide on this site explains them.
Seeds are not portable across GPUs or precisions; expect small differences between your PC and the cloud with the same seed.
When a model page lists a shift or a sigma value, it matters more than the sampler. Set it first.

Compare the saved settings line; expect small drifts between machines.

Every setting is a flag

PirateDiffusion exposes steps, guidance, sampler, seed, size and shift as flags on the /render command, and remembers your favourites. Run a seed sweep from your phone; unlimited generation on a fixed price.

Essayer PirateDiffusion

Questions

Which sampler is best?

There is no best; Euler and DPM++ 2M are the dependable ones. Differences between modern samplers at the same step count are small compared with the model and the prompt.

Is CFG 7 a universal default?

It was for SD 1.5. SDXL likes 5 to 6, Flux dev wants 1 with guidance 3.5, distilled models 1 to 2. Universal defaults are what burn Flux images.

What does the seed actually encode?

Only the random starting noise. It carries no information about the prompt; a seed that gave a good image with one prompt is not special for another.

Why do video models have two CFG values?

Wan 2.2 and others run a high-noise expert for the first half of the steps and a low-noise expert for the second; each has its own CFG and often its own LoRA.