Nodes/Comfyui-DLN-RTX-Tile-Refiner/DLN RTX Tile Refiner
ComfyUI Node

DLN RTX Tile Refiner

Splits an image into a grid of smaller tiles, applies KSampler with a lower denoise to each, and merges them back. Best for upscaling or adding high-res details.

By rauldlnx10·Created 5 months ago·Updated 4 months ago· 2
DLN RTX Tile Refiner
  • image
  • model
  • positive
  • negative
  • vae
  • mask
  • image
seed0
steps20
cfg8.0
sampler_name
scheduler
denoise0.25
grid_divisions
overlap32
modeRefiner (Same Size)
rtx_upscale_factor1.00
rtx_qualityHIGH
blend_modeSmoothstep
sharpening0.00
color_match_strength0.00
detail_inject0.00
context_padding0
CategoryDLN/Image

Inputs (22)

NameTypeDefaultDescription
imageIMAGE
modelMODEL
positiveCONDITIONING
negativeCONDITIONING
vaeVAE
seedINT00–18446744073709550000
stepsINT201–10000
cfgFLOAT8.00–100
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
denoiseFLOAT0.250–1
grid_divisionsCOMBO5 options: 2x2, 4x4, 8x8, 16x16, 32x32
overlapINT320–512Overlap between tiles in pixels to prevent visible seams. A smart minimum is enforced automatically.
modeCOMBORefiner (Same Size)Refiner keeps the original resolution but adds details. Upscale increases the final image resolution based on the RTX factor.
rtx_upscale_factorFLOAT1.001–4Upscale factor using Nvidia RTX before refining. 1.0 disables it (requires nvvfx).
rtx_qualityCOMBOHIGH4 options: LOW, MEDIUM, HIGH, ULTRA
blend_modeCOMBOSmoothstepBlending algorithm for tile transitions. Smoothstep and Gaussian are much smoother than Linear.
sharpeningFLOAT0.000–1Unsharp mask strength applied to the final result. 0=disabled. 0.3-0.5 recommended for crisp details.
color_match_strengthFLOAT0.000–1Match refined tile colors to the original to prevent color drift between tiles. 0.7-0.9 recommended.
detail_injectFLOAT0.000–0.5Inject noise into the latent before sampling to encourage micro-detail generation without increasing denoise. 0.05-0.15 recommended.
context_paddingINT00–128Extra pixels extracted around each tile for KSampler context. Discarded after processing. 32-64 recommended for better tile coherence.
maskoptMASKOptional mask to only refine specific areas. Tiles not covered by the mask will be skipped.

Outputs (1)

NameTypeDescription
imageIMAGEThe refined and merged image