PolyExponential Sigma Adder

The README labels this one "Highres Fix Flux," and that's the honest pitch: a sigma-curve tweak meant for img2img and high-res-fix passes on Flux-class models. Where ModelSamplingFlux gives you a resolution-dependent shift that behaves differently at 512² than at 2048², this node adds a resolution-independent curve adjustment to your sigmas. Same result at any size - which is exactly what you want when your upscale pass runs at inconsistent resolutions.

Mechanically it's a SIGMAS-to-SIGMAS node: feed it the sigma schedule from a scheduler, and it reshapes the curve using two knobs, then hands back a new SIGMAS for your sampler. The math (visible in the source) builds a polyexponential ramp with exponent curves_rigidity, then linearly blends your original sigmas with that curve based on polyexponential_multipier. Roughly: rigidity shapes the curve's bend, multiplier decides how much of the new curve gets mixed in, and the sign flips whether you push the curve up or down.

The three inputs

• sigmas - the schedule to reshape, from a scheduler node.
• curves_rigidity - the curve's stiffness, default 1.5, range −10 to 10. Higher = more aggressive bend; this is the "rigidity" the name promises.
• polyexponential_multipier - how strongly the polyexponential curve is applied, default 0 (off), −5 to 5. The tooltip calls it "finetuning sigmas." A negative multiplier applies a negative poly-exponential curve, which is how you'd drag the noise profile in the opposite direction.

Output is SIGMAS, wired into a KSampler/sampler-custom sigma input, replacing whatever your scheduler produced.

Install and honest expectations

Manager → search TBG Takeaways → install, restart:

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

Under TBG/Takeaways. No deps, no downloads.

The straight talk: this is a fiddly knob with opaque tooltips ("finetuning sigmas" tells you what it does, not what value you want), and it's from a one-dev experimental pack, so plan on A/B testing. Start with polyexponential_multipier near zero and small curves_rigidity changes, and compare at your actual upscale resolution - the whole selling point is that the curve stays consistent across sizes, so test it at two different resolutions and see if your texture retention matches. If you've never touched sigma schedules, this is not the first node you should reach for; but if you're fighting resolution-dependent sharpening differences in your upscale chain, it's a genuinely different tool than the stock ModelSampling nodes.