Nodes/Arctenoxs-Essentials_ComfyUI/KSampler Refiner (Arctenox's Essentials)
ComfyUI Node

KSampler Refiner (Arctenox's Essentials)

Second-pass refiner / hi-res fix node. Typical hi-res fix workflow: 1. KSampler (Arctenox's Essentials) at 512x768, denoise=1.0 --> LATENT 2. KSampler Refiner, upscale_method=bilinear, scale_factor=1.5, denoise=0.45, same seed/sonar --> refined LATENT + IMAGE at 768x1152 Denoise guide: 0.30-0.50 Subtle hi-res fix (adds crisp detail, keeps composition) 0.50-0.70 Moderate rework (allows layout shifts) 0.70-1.00 Heavy rework (treats it almost like a fresh generation)

By Arctenox·Created 8 months ago·Updated about a month ago· 1
KSampler Refiner (Arctenox's Essentials)
  • model
  • positive
  • negative
  • latent
  • vae
  • model
  • positive
  • negative
  • latent
  • vae
  • image
upscale_methodbilinear
scale_factor1.50
target_width0
target_height0
seed0
sonar0
steps20
cfg4.00
sampler_name
scheduler
denoise0.45
vae_decodetrue
CategoryArctenox Essentials/Sampling

Inputs (17)

NameTypeDefaultDescription
modelMODEL
positiveCONDITIONING
negativeCONDITIONING
latentLATENT
upscale_methodCOMBObilinearLatent upscale interpolation method. 'None' skips upscaling.
scale_factorFLOAT1.500.25–8Upscale multiplier. Ignored if target_width/height > 0.
target_widthINT00–8192Target pixel width after upscale. 0 = use scale_factor instead.
target_heightINT00–8192Target pixel height after upscale. 0 = use scale_factor instead.
seedINT0-9223372036854776000–18446744073709550000Seed for refiner pass. Use the same seed as the main KSampler for consistency.
sonarINT0-18446744073709550000–18446744073709550000Golden-ratio sonar offset. Match to main KSampler for coherent variation.
stepsINT201–10000Refiner sampling steps.
cfgFLOAT4.000–100CFG scale for refiner pass.
sampler_nameCOMBO44 options: euler, euler_cfg_pp, euler_ancestral, euler_ancestral_cfg_pp, heun, heunpp2, +38
schedulerCOMBO9 options: simple, sgm_uniform, karras, exponential, ddim_uniform, beta, +3
denoiseFLOAT0.450–1How much the refiner changes the latent. 0.3–0.5 = subtle hi-res fix (adds detail, preserves composition) 0.5–0.7 = moderate rework 0.7–1.0 = heavy rework (similar to full generation)
vae_decodeCOMBOtrueDecode the refined latent to an IMAGE inside this node.
vaeoptVAE

Outputs (6)

NameTypeDescription
modelMODEL
positiveCONDITIONING
negativeCONDITIONING
latentLATENT
vaeVAE
imageIMAGE