KSampler (Advanced)

The KSampler (Advanced) node provides fine-grained and modular control of the denoising process in ComfyUI, letting advanced users customize every aspect of the sampling loop for creative experiments, iterative refinement, and special workflows. It expands on the basic KSampler by enabling partial denoising, noise injection, multi-stage workflows, and beyond.

Overview

KSampler (Advanced) supports advanced use cases such as partial denoising cycles, handover between samplers, and multi-stage inpainting by enabling you to skip noise addition, control denoising boundaries (start_at_step, end_at_step), or even return intermediates for further processing. The node is essential for users building workflows that require more than “start-to-end” diffusion or that need reproducible, highly-customizable image generation.

Visual Example

KSampler (Advanced) ComfyUI node
Figure 1 - KSampler (Advanced) ComfyUI node

Official Documentation Link

https://blenderneko.github.io/ComfyUI-docs/Core%20Nodes/Sampling/KSampler%20Advanced/

Inputs

Parameter Data Type Input Method Default
model Object Model loader node output —
latent_image Tensor Node input/connection —
positive Embedding Prompt conditioning input —
negative Embedding Prompt conditioning input —
steps Integer Numeric field/slider 20
cfg Float Numeric field/slider 7.0
sampler_name String/Option Dropdown (e.g., "euler", "dpmpp_2m", etc.) euler
scheduler String/Option Dropdown (auto, or specific scheduler) automatic
add_noise Boolean Checkbox True
return_with_leftover_noise Boolean Checkbox False
seed Integer Numeric field Random
start_at_step Integer Numeric field 0
end_at_step Integer Numeric field steps

Outputs

Output Name Data Type Description
latent Tensor Sampled latent tensor after denoising (partial or full) ready for decoding

Usage Instructions

Add a KSampler (Advanced) node to your workflow canvas. Connect your model, latent_image, positive, and negative conditioning nodes as appropriate. Set the desired values for steps, cfg, sampler_name, and scheduler. Use add_noise to toggle noise addition (disable for partial denoising or refine cycles). Tune start_at_step and end_at_step for iterative or partially completed denoising. Enable return_with_leftover_noise if you want intermediary outputs for progressive workflows. Use the output latent as the basis for further refinement or final decode to image using VAE Decode.

Advanced Usage

Chain several KSampler (Advanced) nodes by handing off leftover noise/tensors for iterative, multi-stage denoising or inpainting. Perform partial denoise + add-and-refine cycles for repairing, upscaling, or region-based edits. Benchmark and experiment using a variety of samplers and schedulers, switching mid-workflow using advanced routing nodes. Pair with custom control nodes or script-based testing nodes for research and ablation studies.

Example JSON for API or Workflow Export

{
  "id": "ksampler_adv_1",
  "type": "KSamplerAdvanced",
  "inputs": {
    "model": "@load_diffusion_model_1",
    "latent_image": "@emptysd3latentimage_1",
    "positive": "@conditioning_positive_1",
    "negative": "@conditioning_negative_1",
    "steps": 40,
    "cfg": 7.0,
    "sampler_name": "dpmpp_2m",
    "scheduler": "automatic",
    "add_noise": true,
    "return_with_leftover_noise": false,
    "seed": 1337,
    "start_at_step": 10,
    "end_at_step": 40
  }
}

Tips

  • Only enable add_noise for initial denoise cycles; disable for iterative refinements.
  • Set start_at_step and return_with_leftover_noise for multi-stage relaxed or specialized pipelines.
  • Save seeds and all parameters for reproducibility in collaborative or research settings.
  • If you want only partial denoising, set start/end at less than full step range and chain accordingly.
  • For speed, reduce steps; for maximal quality, tune cfg/sampler/scheduler combinations.

How It Works (Technical)

KSampler (Advanced) executes a configurable subset of the standard denoising schedule. It may skip noise addition or start/stop mid-way, allowing for nonstandard workflows and partial processing. It supports outputting both the current latent and any leftover noise, making it ideal for progressive, iterative, or experimental workflows where composable pipeline steps are required.

Github Alternatives

  • ComfyUI-TripleKSampler – Provides simple, advanced, and alternative KSampler nodes for advanced and chained sampling workflows.
  • Efficiency Nodes for ComfyUI – Includes KSampler Adv. (Efficient), supporting live previews, partial denoise, and flexible workflow design.
  • KSampler Tester/Loop – Lets you systematically test combinations of sampler parameters in batch and looped workflows, featuring advanced and custom sampler support.

Videcool workflows

The KSampler (Advanced) node is used in the following Videcool workflows:

FAQ

1. How is KSampler (Advanced) different from basic KSampler?
It adds settings for fine control: disables/enables noise addition, sets step start/end, returns partial/noise for advanced chaining.

2. How do I perform multi-stage denoising?
Chain multiple KSampler (Advanced) nodes, adjusting start/end steps and input latent/tensor at each stage.

3. Can I use all the same samplers/schedulers?
Yes; this node exposes and enhances all sampler/scheduler options from core and plug-in extensions.

Common Mistakes and Troubleshooting

Common mistakes include forgetting to disable add_noise for refinement cycles, leading to repeated unnecessary denoising. Mismatched latent and model inputs can cause errors; always verify correct node chaining. Leaving start/end steps on defaults (0/full) when partials are desired is problematic; always tune explicitly. Memory overloads on long chains can be addressed by reducing batch, image size, or cleaning up in multi-stage workflows. Seed management is important—ensure seeds propagate for reproducibility if comparing results across runs.

Conclusion

KSampler (Advanced) is a power-user’s tool for building complex, robust, and fully-customized image generation pipelines in ComfyUI. With detailed step, chain, and noise controls, it is essential for specialized workflows, research, and creative AI experiments.

More information