TBG_FluxKontextStabilizer

Here's a problem tiled-upscaling people know well: when you refine a large image tile by tile, every tile runs its own sampling pass, and unless the sigma schedule is identical across tiles, the model interprets each pass slightly differently - the same prompt, same seed, slightly different composition. Those differences read as seams, drift, and position shifts. This node is TBG's fix: it takes whatever SIGMAS your scheduler produced and replaces it with a fixed, hand-tuned 21-step curve, so every tile runs on the exact same schedule and keeps the same spatial interpretation.

It's the "Kontext" in the name - the author built it for his TBG ETUR tiled upscaler to keep the context of reference images pinned through refinement. The mechanism in the source is blunt and effective: it keeps the first six steps of one known-good curve, then filters the incoming sigmas down to values below a threshold and appends them, producing a consistent schedule regardless of what your scheduler originally calculated.

What it does and doesn't do

One input, one output: sigmas → SIGMAS. Drop it between your scheduler and sampler (or wherever your tiled pass computes its schedule) and it normalizes the schedule. Note the author's operating instructions in the README, because they're the actual recipe: stay with the euler sampler and beta scheduler, 16–30 steps, and add something like "Repair and enhance this photo" to your prompt. That's not optional flavor - the fixed curve is tuned for that sampler/scheduler window, and other combos will behave differently.

Install

Manager → search TBG Takeaways → install, restart:

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

Under TBG/Takeaways after restart. No deps, no downloads.

Honest take: this is the most niche node in the pack. It's essentially an internal component of the author's upscaler, released standalone "for anyone who enjoys experimenting." If you're building a multi-tile refinement workflow with Flux-class models and fighting tile-to-tile drift, it's worth a test - it's one node, zero cost to try, and the fixed-schedule insight is real. If you run single-pass generation, skip it entirely; it does nothing for you. And like everything in this pack, it's a one-dev experimental tool: no docs beyond the README's sampler/scheduler advice, so treat that advice as the ground truth for tuning.