ComfyUI Node

Flow Matching Stage

A ComfyUI node in latent/upscaling with 21 inputs and 6 outputs.

By ttulttul·Created 9 months ago·Updated 6 months ago· 62
Flow Matching Stage
  • model
  • positive
  • negative
  • latent
  • latent
  • presampler_latent
  • next_seed
  • model
  • positive
  • negative
seed0
steps16
cfg4.5
sampler_name
scheduler
scale_factor1.00
noise_ratio0.00
skip_blend0.50
denoise1.00
upscale_methodbicubic
enable_dilated_samplingdisable
reduce_memory_useenable
dilated_downscale2.00
dilated_blend0.25
dilated_min_steps1
dilated_seed_modederive
dilated_denoise0.50
Categorylatent/upscaling

Inputs (21)

NameTypeDefaultDescription
modelMODELFlow-matching diffusion model to drive refinement.
positiveCONDITIONINGPositive conditioning for CFG.
negativeCONDITIONINGNegative conditioning for CFG.
latentLATENTLatent to upscale and refine in this single stage.
seedINT00–18446744073709550000Seed controlling re-noising for this stage (and typically the sampler noise).
stepsINT161–256Denoising steps for this stage.
cfgFLOAT4.50–20Classifier Free Guidance strength.
sampler_nameCOMBOSampler backend leveraged during refinement.
schedulerCOMBONoise schedule applied during denoising.
scale_factorFLOAT1.000.1–8Spatial scale factor applied to the latent grid for this stage.
noise_ratioFLOAT0.000–1Flow-style re-noise amount (0 = keep latent, 1 = replace with pure noise).
skip_blendFLOAT0.500–1Skip blend weight (0 = all denoised, 1 = all pre-sampler latent).
denoiseFLOAT1.000–1Denoising strength supplied to the sampler.
upscale_methodCOMBObicubicResampling kernel for spatial upscaling. NOTE: ComfyUI's `lanczos` path uses PIL and is unsafe for LATENT tensors; this node will fall back to `bicubic` if selected.
enable_dilated_samplingoptCOMBOdisableOptionally run a dilated refinement pass for global coherence (experimental).
reduce_memory_useoptCOMBOenableEnable to reduce VRAM use by avoiding extra tensor clones where possible.
dilated_downscaleoptFLOAT2.001–4Factor used when downscaling for the dilated pass (>=1.0).
dilated_blendoptFLOAT0.250–1Blend weight of the dilated refinement result (frequency-domain blend).
dilated_min_stepsoptINT11–256Minimum sampling steps for dilated refinement. Actual steps = max(min_steps, steps // 2). Increase for lightning models to ensure adequate sampling (e.g., 4).
dilated_seed_modeoptCOMBOderiveSeed mode for dilated sampling. 'derive' adds 10,000 to base seed, 'same' uses identical seed.
dilated_denoiseoptFLOAT0.500–1Denoising strength for dilated sampling. Lower values preserve more spatial structure from the original.

Outputs (6)

NameTypeDescription
latentLATENT
presampler_latentLATENT
next_seedINT
modelMODEL
positiveCONDITIONING
negativeCONDITIONING