.ai

les meilleurs modèles d'IA open source

Rechercher

Nœud KSampler dans ComfyUI

KSampler runs the denoising loop: it takes the model, prompts and a latent and produces the finished latent image, controlled by seed, steps, cfg, sampler, scheduler and denoise.

Pack de nœuds : ComfyUI core

Catégorie : Échantillonnage

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

Ce que fait KSampler

KSampler is where the image is actually generated. Starting from the input latent (random noise from Empty Latent Image, or an encoded picture for img2img), it asks the model to remove noise for the chosen number of steps, pulling toward the positive prompt and away from the negative one by the cfg amount. The result is a LATENT that VAE Decode turns into an image.

sampler_name selects the numerical method (euler, euler_ancestral, dpmpp_2m, dpmpp_2m_sde, dpmpp_3m_sde, uni_pc, lcm, ...) and scheduler the noise curve (normal, karras, exponential, sgm_uniform, simple, beta). denoise controls how much of the input latent is replaced: 1.0 for text-to-image, 0.3-0.7 for img2img and hires passes. KSampler (Advanced) exposes start and end steps for multi-stage sampling.

Entrées

Nom	Type	Rôle
model	MODEL	The denoising model, after any LoRA or ControlNet patches.
positive	CONDITIONING	What you want (from CLIP Text Encode, possibly through ControlNet Apply).
negative	CONDITIONING	What you do not want.
latent_image	LATENT	The starting latent: empty noise or an encoded image.
seed	INT	Noise seed. Same seed and settings reproduce the same image. control_after_generate sets fixed, increment, decrement or randomize.
steps	INT	Number of denoising steps. 20-30 for most SD models, 4-8 for Lightning, Turbo and LCM models.
cfg	FLOAT	Classifier-free guidance scale. 5-8 for SD 1.5 and SDXL, 1.0 for Flux and distilled models.
sampler_name	COMBO	The sampling algorithm.
scheduler	COMBO	The noise schedule.
denoise	FLOAT	Fraction of the noise to remove. 1.0 = start from scratch; lower keeps more of the input image.

Sorties

Nom	Type	Rôle
LATENT	LATENT	The sampled latent for VAE Decode or a second sampler.

Oubliez l'installation : des workflows ComfyUI préinstallés sur le cloud

BitVector exécute des workflows ComfyUI prêts à l'emploi sur ses propres GPU. Rien à installer, aucun nœud manquant, aucune case rouge. Ouvrez-le sur votre téléphone ou votre portable et générez en une minute.

Essayer BitVector

Comment utiliser KSampler

Connect MODEL, both CONDITIONINGs and a LATENT.
Set steps 25, cfg 7, dpmpp_2m with karras for SDXL; euler with simple, cfg 1 for Flux.
Set denoise to 1.0 for a fresh image.
Connect LATENT to VAE Decode and then Save Image.
Queue, then fix the seed when you want to iterate on the prompt.

Réglages et conseils

dpmpp_2m + karras is the all-round choice; euler_ancestral adds variety but never converges; dpmpp_2m_sde gives extra detail at the cost of determinism across GPUs.
Lightning, Hyper and Turbo checkpoints need their own settings: 4-8 steps, cfg 1-2, sgm_uniform or simple.
For a hires pass, feed the first latent through Upscale Latent By and a second KSampler with denoise 0.4-0.55.
Hold Ctrl+Shift+V to paste a node with its connections; useful when duplicating samplers.
A fixed seed with the same settings is reproducible on the same GPU; different cards can differ slightly.

Dépannage de KSampler

Noise, static or a blurry blob instead of an image

Pourquoi cela arrive

The settings do not fit the model: cfg 7 on a distilled model, 4 steps on a normal model, a Flux model without Flux Guidance, or denoise far below 1 on an empty latent.

Comment le corriger

Match the model page: normal SD = 20-30 steps, cfg 5-8; Lightning/Turbo = 4-8 steps, cfg 1-2; Flux = cfg 1, guidance 3.5, 20+ steps. Use denoise 1.0 when the latent is empty.

Deep-fried, oversaturated, high-contrast output

Pourquoi cela arrive

cfg is too high for the model or the LoRA stack, or an ancestral sampler with many steps.

Comment le corriger

Lower cfg to 4-6, reduce LoRA strengths, or switch from euler_ancestral to dpmpp_2m.

torch.OutOfMemoryError: CUDA out of memory during sampling

Pourquoi cela arrive

The latent is too large for the card at this batch size, or another model (ControlNet, upscaler) sits in VRAM at the same time.

Comment le corriger

Lower the resolution or batch_size, add --lowvram, use fp8 weights, or split the hires pass into a second queue so the first model is unloaded.

Error: mat1 and mat2 shapes cannot be multiplied

Pourquoi cela arrive

The conditioning comes from a text encoder of a different family than the model (SD 1.5 CLIP into an SDXL model, or vice versa).

Comment le corriger

Wire the CLIP from the same checkpoint that provides the MODEL.

Images change every run although the seed is fixed

Pourquoi cela arrive

control_after_generate is set to randomize, or an SDE / ancestral sampler is in use with a non-deterministic backend.

Comment le corriger

Set control_after_generate to fixed and prefer dpmpp_2m or euler for exact repeatability.

The same image repeats across prompts (seed stuck)

Pourquoi cela arrive

control_after_generate is fixed and the noise dominates at low cfg or low steps.

Comment le corriger

Set it to randomize or increment and raise cfg slightly.

Lancez ce workflow depuis votre téléphone

Chaque workflow de cette page est préinstallé sur BitVector, modèles et nœuds personnalisés déjà en place. Choisissez-en un, tapez un prompt, c'est fait. Zéro configuration, rien à télécharger.

Essayer BitVector

Questions sur KSampler

What is the difference between KSampler and KSampler (Advanced)?

Advanced adds add_noise, start_at_step, end_at_step and return_with_leftover_noise, which lets you split sampling across two models, for example SDXL base then refiner.

Which sampler is fastest?

euler and dpmpp_2m take one model call per step; the _2s and _3m variants take more. LCM and Lightning models are fast because they need few steps, not because of the sampler.

What does denoise do with an empty latent?

Anything below 1.0 leaves part of the original noise in the output and produces fog. Keep 1.0 for text-to-image.

Nœuds liés

KSampler (Advanced)

ComfyUI core

KSampler (Advanced) is KSampler with step-range control: add_noise, start_at_step, end_at_step and return_with_leftover_noise, used for base + refiner chains and two-model sampling.

Empty Latent Image

ComfyUI core

Empty Latent Image creates the blank latent canvas (width, height, batch_size) that text-to-image sampling starts from; dimensions must be multiples of 8 and match the model family.

VAE Decode

ComfyUI core

VAE Decode converts the sampled LATENT into a pixel IMAGE with the VAE; VAE Decode (Tiled) does the same in tiles for very large images.

CLIP Text Encode (Prompt)

ComfyUI core

CLIP Text Encode turns a text prompt into CONDITIONING using the model text encoder. One node holds the positive prompt, a second one the negative prompt.

Upscale Latent and Upscale Latent By

ComfyUI core

Upscale Latent resizes a latent to a new size (or by a factor) so a second KSampler can add detail at higher resolution: the classic hires fix.

Marre de réparer des nœuds ? Laissez faire le cloud

BitVector garde des centaines de workflows ComfyUI installés, à jour et testés sur des GPU cloud rapides. Pas de Python, pas d'erreurs CUDA, pas de limite de VRAM. Depuis n'importe quel navigateur.

Essayer BitVector