LogSigmaSamplerNode

Here's a trick that's quietly popular in the Flux detail crowd: the model has an expected noise curve baked into training, and if you hand it a slightly different sigma than the scheduler says, you can push it to add detail, sharpen, or soften without touching prompts or steps. The community knows this family as Detail Daemon, Lying Sigmas, or "multiply sigmas" - a real, frequently-discussed set of techniques, not a theoretical one. This node is TBG's take on it.

It works by wrapping an existing sampler. You give it a sampler, a LogSigmas curve, and a lerpfactor, and it returns a new SAMPLER that lies to the model about the noise level during part of the run: instead of passing the true sigma, it linearly interpolates between the true sigma and the value from your LogSigmas curve, weighted by lerpfactor. Negative lerpfactor (default −0.05) reduces the sigma the model sees, which nudges it toward finer, more defined output - the tooltip puts it plainly: −0.05 means sigma reduced by 5%.

What you actually wire up

• sampler - any SAMPLER object, e.g. from a SamplerCustom/KSampler setup.
• LogSigmas - a SIGMAS tensor, your custom noise curve. This is the curve the node lerps toward, so it's where the personality comes from. Start with a scheduler's sigmas and reshape, or hand it a curve from another sigma node.
• lerpfactor - how hard to push. −0.05 is gentle; more negative is stronger.
• start_percent / end_percent - the window of the run where the lie is active, as a fraction (defaults 0.1 to 0.9). Outside it, the model gets the truth. This is what keeps the effect surgical instead of destabilizing the whole image.

The output is a SAMPLER, and you feed it into a KSampler/sampler-custom node in place of your original sampler. If you see the effect hit everything and smear the composition, narrow the window.

Install

Manager → search TBG Takeaways → install, restart:

cd ComfyUI/custom_nodes
git clone https://github.com/Ltamann/ComfyUI-TBG-Takeaways

Find it under TBG/Takeaways. No extra deps, no model downloads.

Some honest guidance. This is the "precise, surgical" sibling of the pair - if you're already comfortable with Detail Daemon's workflow, the mechanics will feel familiar (the README credits that family directly). Start with lerpfactor at −0.05, look at a crop at 100% zoom, and nudge from there. The catch is the same as every sigma trick: your LogSigmas curve has to actually make sense as a noise schedule, and the model's expectations vary. Also expect to hit the print() debugging lines the author left in the code on some versions - harmless console spam, ignore it. If this is your first time with sigma manipulation, expect an afternoon of A/B testing before it clicks.