2026-01-28 00:02:49 -03:00
1.0
2026-01-26 04:23:24 -03:00
2026-01-26 04:49:02 -03:00
2026-01-27 23:06:01 -03:00
2026-01-27 23:07:08 -03:00
2026-01-27 23:05:02 -03:00
2026-01-28 00:02:49 -03:00

KSampler DF (with Refiner & Upscale)

screenshot

A custom ComfyUI node that splits sampling steps between a main sampler and a refiner sampler, with latent upscale and optional model-based upscale.

Changelog

v1.0.2 (2026-01-27)

  • 🔧 IMPROVED: Upscale model now selected directly via dropdown (no external node needed)
  • 🔧 IMPROVED: VAE is now required for image output
  • ✨ NEW: Automatic listing of all upscale models from models/upscale_models/

v1.0.1 (2026-01-27)

  • ✨ NEW: Latent upscale between main and refiner phases
  • ✨ NEW: Separate CFG control for refiner phase (cfg_refiner)
  • ✨ NEW: Optional model-based upscale after refinement
  • ✨ NEW: Dual output: LATENT and IMAGE

v1.0.0 (Initial Release)

  • Basic KSampler with refiner functionality
  • Split steps between main and refiner samplers
  • Separate denoise control for each phase

Features

  • Split Steps: Divide total steps between main sampler and refiner sampler
  • Dual Sampler Support: Use different samplers/schedulers for main and refiner phases
  • Independent CFG Control: Separate CFG values for main and refiner phases
  • Independent Denoise Control: Separate denoise parameters for main and refiner phases
  • Latent Upscale: Optional upscale of latent between main and refiner phases
  • Model Upscale: Built-in dropdown to select upscale models (ESRGAN, RealESRGAN, etc.)

Installation

cd ComfyUI/custom_nodes
git clone https://github.com/YOUR_USERNAME/ksamplerwithrefiner.git

Restart ComfyUI after installation.

Method 2: Manual Installation

  1. Download this repository as a ZIP file
  2. Extract to ComfyUI/custom_nodes/ksamplerwithrefiner/
  3. Restart ComfyUI

Usage

Find the node under: sampling → KSampler DF (with Refiner & Upscale)

Parameters

Main Sampling

Parameter Description
model The diffusion model
positive/negative Conditioning (prompts)
latent_image Input latent
seed Random seed
steps Total steps (main + refiner)
cfg CFG scale for main phase
sampler_name Sampler for main phase
scheduler Scheduler for main phase
denoise Denoise for main phase

Latent Upscale

Parameter Description
latent_upscale_enabled Enable/disable
latent_upscale_method nearest-exact, bilinear, area, bicubic, bislerp
latent_upscale_factor Scale factor (e.g., 1.5)

Refiner

Parameter Description
refiner_at_step Step to switch to refiner
sampler_refiner Sampler for refiner phase
scheduler_refiner Scheduler for refiner phase
cfg_refiner CFG scale for refiner phase
denoise_refiner Denoise for refiner (0.0-1.0)

Output & Upscale

Parameter Description
vae VAE for decoding to image
upscale_model_name Select upscale model or "none" to disable

Outputs

Output Description
LATENT Denoised latent
IMAGE Decoded/upscaled image

Example

  • steps: 20, refiner_at_step: 12

Result: 12 main steps → (latent upscale) → 8 refiner steps → VAE decode → (model upscale)

Upscale Models

Place upscale models in ComfyUI/models/upscale_models/. They will appear automatically in the dropdown.

Supported models: ESRGAN, RealESRGAN, SwinIR, and other Spandrel-compatible models.

Requirements

  • ComfyUI (latest version)
  • No additional dependencies

License

MIT License

S
Description
No description provided
Readme MIT
129 KiB
Languages
Python 100%