Skip to main content
Sampling nodes generate images by iteratively denoising latent images using a diffusion model guided by conditioning.

​
Basic Sampling

​
KSampler

The main node for generating images through the sampling process. Category: sampling
​
MODEL
required
The diffusion model used for denoising the latent
​
int
default:"0"
Random seed for noise generation. Range: 0 to 18446744073709551615
​
int
default:"20"
Number of denoising steps. More steps = higher quality but slower. Range: 1 to 10000
​
float
default:"8.0"
Classifier-Free Guidance scale. Higher values follow the prompt more closely. Range: 0.0 to 100.0
​
string
required
The sampling algorithm to use (e.g., euler, euler_a, dpmpp_2m, dpmpp_2m_sde, heun, dpm2, etc.)
​
string
required
Controls the noise schedule (e.g., normal, karras, exponential, sgm_uniform, simple, ddim_uniform)
​
CONDITIONING
required
The positive prompt conditioning (what you want to generate)
​
CONDITIONING
required
The negative prompt conditioning (what you want to avoid)
​
LATENT
required
The latent to denoise (from EmptyLatentImage for txt2img, or VAEEncode for img2img)
​
float
default:"1.0"
Amount of denoising. 1.0 = full denoising, lower values preserve more of the input. Range: 0.0 to 1.0
Returns:
  • LATENT: The denoised latent image
Description: Uses the provided model, positive and negative conditioning to denoise the latent image. This is the core image generation node. Common Sampler Algorithms:
  • euler: Fast, good quality, deterministic
  • euler_a: Ancestral euler, more creative but non-deterministic
  • dpmpp_2m: DPM++ 2M, high quality with fewer steps
  • dpmpp_2m_sde: DPM++ 2M SDE, higher quality but slower
  • dpmpp_3m_sde: DPM++ 3M SDE, very high quality
  • heun: High quality but slow (2x function evaluations)
  • dpm_fast: Fast sampling
  • ddim: Classic DDIM sampler
  • uni_pc: UniPC sampler
  • lcm: Latent Consistency Model sampler (requires LCM LoRA)
Common Schedulers:
  • normal: Standard linear schedule
  • karras: Karras schedule, often better quality
  • exponential: Exponential decay schedule
  • sgm_uniform: Stable diffusion 3 schedule
  • simple: Simple schedule
  • ddim_uniform: Uniform schedule for DDIM
CFG Scale Guidelines:
  • 1.0-3.0: Very loose adherence, highly creative
  • 4.0-7.0: Balanced creativity and prompt following
  • 7.0-12.0: Strong prompt adherence
  • 12.0+: Very strict, may reduce quality

​
Advanced Sampling

​
KSamplerAdvanced

Advanced sampler with fine control over the sampling process. Category: sampling
​
MODEL
required
The diffusion model
​
string
default:"enable"
Whether to add noise at the start. Options: enable, disable
​
int
default:"0"
Random seed for noise. Range: 0 to 18446744073709551615
​
int
default:"20"
Total number of sampling steps. Range: 1 to 10000
​
float
default:"8.0"
Classifier-Free Guidance scale. Range: 0.0 to 100.0
​
string
required
The sampling algorithm
​
string
required
The noise scheduler
​
CONDITIONING
required
Positive conditioning
​
CONDITIONING
required
Negative conditioning
​
LATENT
required
Input latent
​
int
default:"0"
Which step to start sampling at. Range: 0 to 10000
​
int
default:"10000"
Which step to end sampling at. Range: 0 to 10000
​
string
default:"disable"
Whether to return without final denoising. Options: disable, enable
Returns:
  • LATENT: The denoised latent
Description: Provides advanced control over the sampling process. Useful for:
  • Multi-stage sampling (different prompts at different steps)
  • Noise control for advanced workflows
  • Partial sampling for refinement
Use Cases: Two-Stage Sampling:
  1. First KSamplerAdvanced: start_at_step=0, end_at_step=15, return_with_leftover_noise=enable
  2. Second KSamplerAdvanced: start_at_step=15, end_at_step=20, different prompt
Refining:
  • Use add_noise=disable to refine an existing latent without adding noise

​
Sampling Parameters Guide

​
Steps

More steps generally produce better quality but take longer:
  • 15-20: Fast, good for testing
  • 20-30: Standard quality
  • 30-50: High quality
  • 50+: Diminishing returns
Some samplers (like dpmpp_2m) produce good results with fewer steps.

​
CFG (Classifier-Free Guidance)

Controls how closely the output matches the prompt:
  • Too low (< 3): Image may not match prompt
  • Optimal (7-10): Good balance
  • Too high (> 15): May introduce artifacts
Different models have different optimal CFG ranges. SD1.5 typically uses 7-12, while SDXL often works well at 4-8.

​
Denoise Strength

Controls how much the input is changed:
  • 1.0: Complete regeneration (txt2img)
  • 0.7-0.9: Strong changes while preserving composition
  • 0.4-0.6: Moderate changes, refinement
  • 0.1-0.3: Subtle changes, details only
  • 0.0: No change
Used for img2img workflows where you want to modify an existing image.

​
Seed

Controls randomness:
  • Same seed with same settings = same image
  • Different seeds = variations
  • control_after_generate setting:
    • fixed: Keep same seed
    • increment: Add 1 to seed each generation
    • decrement: Subtract 1 each generation
    • randomize: Use random seed each time

​
Workflow Examples

​
Text to Image (txt2img)

  1. EmptyLatentImage → creates blank latent
  2. CLIPTextEncode (positive) → encodes prompt
  3. CLIPTextEncode (negative) → encodes negative prompt
  4. KSampler → generates image
  5. VAEDecode → converts to pixels

​
Image to Image (img2img)

  1. LoadImage → loads input image
  2. VAEEncode → converts to latent
  3. KSampler with denoise < 1.0 → modifies image
  4. VAEDecode → converts back to pixels

​
Inpainting

  1. LoadImage → loads image
  2. LoadImageMask → loads mask
  3. VAEEncodeForInpaint → prepares latent with mask
  4. KSampler → generates in masked area
  5. VAEDecode → final image

​
Upscaling (Latent Upscale)

  1. Generate base image at low resolution
  2. LatentUpscale → upscale latent
  3. KSampler with denoise=0.4-0.6 → refine details
  4. VAEDecode → high-res image