🎚️ Denoising Strength

If you've used img2img, you know the denoising strength slider: 0.3 keeps the composition, 0.7 basically regenerates, and somewhere in between is where the magic lives. The KSampler in ComfyUI has a denoise field that does this for you - unless you're feeding the sampler a raw sigma schedule, in which case there's no denoise knob at all. DenoisingStrengthModifier is that knob, applied directly to a SIGMAS tensor. It's the bridge between "I want img2img-style control" and "I'm driving sampling from a custom sigma schedule."

The inputs

• sigmas (required, type SIGMAS) - the noise schedule you want to modify, typically from this pack's Custom Graph Sigma, Sigma Joiner, or Preset Sigma Generator.

• denoising_strength (FLOAT, 0–1, default 1) - exactly what it says on the tin. 1.0 = leave the schedule alone (full txt2img-style denoise); lower = less of the schedule is actually used.

• mode (dropdown) - how the strength is applied, and the part worth understanding:

  • truncate (default) - cuts the schedule off from the high-noise end. This is the classic img2img behavior: you start from a partially noised image, so you skip the high-noise steps entirely. It's the mode that matches what the KSampler's denoise slider does.
  • scale - multiplies every sigma value by the strength. The whole schedule shrinks, so the overall noise level is lower. That's not quite the same thing as fewer steps; it's a quieter run.
  • interpolate - tapers the schedule smoothly down to zero, using weights from the strength down to 0. A gentler, more continuous reduction than truncation.

The outputs

• modified_sigmas (SIGMAS) - your new schedule. Wire this into wherever the original sigmas was going.
• effective_steps (INT) - how many steps the modified schedule actually contains. In truncate mode this shrinks as you lower the strength, which is genuinely useful for wiring a downstream step count or for understanding why your 20-step workflow suddenly ran 8 steps.

The mental model

A sigma schedule is just the list of noise levels the sampler walks down, largest to smallest. "Denoising strength 0.5" means "only walk the second half of that walk." Truncation does that literally - it removes the top half of the schedule, which is exactly what it means to start from a partially noised image. The other two modes are variations for when you want the schedule reshaped rather than shortened.

Practical setting guide (the numbers the community actually uses, from the KB): ~0.3–0.5 preserves composition while allowing a style shift, 0.7+ is near-reinterpretation, and below ~0.2 you're doing touch-ups - at the extreme low end, some models visibly shift color even at "0%" denoising just from the VAE round-trip, so don't be surprised if a near-zero strength isn't pixel-identical.

Where it plugs in

The output SIGMAS goes into any sampler node that accepts a raw sigma tensor - in stock ComfyUI that's typically SamplerCustom, or another node in this pack's sigma pipeline. If you're not working with custom sigmas yet, this node is solving a problem you don't have; the KSampler's denoise field is the simpler path. This one earns its keep once you've moved to schedule-driven sampling.

Install

Pack: InitialB Util. ComfyUI Manager → search "InitialB Util" → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/benjiyaya/Comfyui_InitialB_Util
cd Comfyui_InitialB_Util
pip install -r requirements.txt

Restart after. No model downloads; requirements are torch/numpy/scipy/Pillow, all bundled with ComfyUI. (The README's clone URL is a yourusername placeholder - use the real repo above or Manager.) It sits in InitialB/sigma/custom with the rest of the schedule tools.