Nodes/TBG’s ComfyUI Development Takeaways/TBG Dual Model KSampler (Inpaint Split Aware)
ComfyUI Node

TBG Dual Model KSampler (Inpaint Split Aware)

A ComfyUI node in TBG/Sampler with 20 inputs and 1 output.

By Ltamann·Created about a year ago·Updated 7 months ago· 19
TBG Dual Model KSampler (Inpaint Split Aware)
  • model_high
  • model_low
  • positive_high
  • negative_high
  • positive_low
  • negative_low
  • latent_image
  • output
noise_seed0
cfg_high1.0
cfg_low1.0
sampler_name
scheduler
steps_high20
steps_low9
denoise1.00
model_crossover_sigma_strength0.50
low_sigma_alignment1.00
inpaint_end0
smoother_sharper0.00
detail_enhancer0.00
CategoryTBG/Sampler

Inputs (20)

NameTypeDefaultDescription
model_highMODELFirst model for high sigma phase. IMPORTANT: Both model_high and model_low must use the same or compatible latent spaces and VAE models. Compatible pairs: Flux + ZImages, Qwen + WAN. Incompatible models will produce corrupted outputs due to latent space mismatch.
model_lowMODELSecond model for low sigma phase. IMPORTANT: Must be latent space compatible with model_high (same VAE, same latent dimensions). Compatible pairs: Flux + ZImages, Qwen + WAN. Using incompatible models will fail or produce artifacts.
noise_seedINT00–18446744073709550000
cfg_highFLOAT1.00–100
cfg_lowFLOAT1.00–100
positive_highCONDITIONING
negative_highCONDITIONING
positive_lowCONDITIONING
negative_lowCONDITIONING
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
steps_highINT201–2048Total steps used to generate the high model's sigma schedule. This is the full step count for a typical generation, not the actual steps executed. The sigma_split_value determines which portion of this schedule is used. Example: FLUX 20 steps with split=0.5 might execute ~10 high-sigma steps.
steps_lowINT91–2048Total steps used to generate the low model's sigma schedule. This is the full step count for a typical generation, not the actual steps executed. The sigma_split_value determines which portion of this schedule is used. Example: Z-image 10 steps with split=0.5 might execute ~5 low-sigma steps.
denoiseFLOAT1.000–1
model_crossover_sigma_strengthFLOAT0.500–1Controls the sigma strength where sampling switches from the high model to the low model.”
low_sigma_alignmentFLOAT1.000–5Shifts the lower sigma values up or down to better synchronize sampling between models. Default is 1, which works best in most cases.
inpaint_endINT0-50–0Step number from the end after which inpainting is skipped. For example, with 20 total steps, setting -10 means inpainting runs only from step 1 to 10.
smoother_sharperFLOAT0.00-1–1Dual-stage adaptive sharpening. At high sigma (early steps), adds structured noise for detail invention. At low sigma (late steps), applies high-pass edge sharpening. Positive values sharpen and add details. Negative values soften and blur. Zero disables sharpening. Higher absolute values create stronger effects.
detail_enhancerFLOAT0.00-1–1Substep evaluation for detail control. Positive values (0.1-1.0): lookahead to next sigma, adds coherent details and refinement, reduces variation. Negative values (-0.1 to -1.0): lookback to previous sigma, adds creative variation and texture complexity. Zero = disabled (single pass, fastest). Performance cost: 2x slower on affected steps.
latent_imageoptLATENT

Outputs (1)

NameTypeDescription
outputLATENT