TBG KSampler Advanced (Inpaint Split Aware)

At a glance this looks like the standard KSamplerAdvanced with extra sliders bolted on, and in a good way. TBG's take on the advanced sampler keeps everything you'd expect - start_at_step, end_at_step, add_noise, return_with_leftover_noise - then adds three genuinely useful controls the stock node doesn't have: an explicit inpainting window, a sharpen/soften stage, and a detail enhancer that changes how the sampler evaluates at each step.

Where people get burned by stock ComfyUI inpainting is timing. The classic KSamplerAdvanced blends the original latent back into the masked area for the whole run, which is often not what you want - you want the model free to invent during the noisy early steps and only pinned to the original in the detail phase, or the reverse. This node gives you inpaint_start and inpaint_end, literally "step numbers where inpainting injection begins/ends." Set inpaint_start higher to delay mask enforcement and let the model improvise; lower inpaint_end to let late steps refine without the mask clamping them. It's the difference between "blend this region" and "redraw this region, then lock it."

What the extra controls do

• smoother_sharper (slider, −1 to 1) - a dual-stage effect: at high sigma it adds structured noise for detail invention, at low sigma it applies high-pass edge sharpening. Positive sharpens, negative softens, zero off. It's adaptive, not a fixed unsharp mask.
• detail_enhancer (slider, −1 to 1) - substep lookahead/lookback. Positive values peek at the next sigma to add coherent detail; negative values look back for creative variation. Zero is disabled, and it's worth knowing it's the fastest setting - the tooltip is honest that this roughly doubles cost on affected steps.
• sampler_state (optional) - a chaining token. Leave empty on the first sampler, connect the output to the next sampler's sampler_state for multi-pass workflows. This is the "split aware" part: the state carries the sigma schedule across passes so chained samplers stay coherent.

Everything else behaves like the stock advanced sampler: cfg defaults to 1 (right for Flux, wrong for SD1.5-style guidance - bump it if your checkpoint expects CFG), denoise, sampler_name, scheduler, and the 44 sampler choices that ship with ComfyUI. Outputs are LATENT plus sampler_state.

Install and setup

Same story as the rest of the pack - ComfyUI Manager, search TBG Takeaways, install, restart. Manual route:

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

Restart and it appears under TBG/Sampler. No Python deps beyond stock ComfyUI, no model files to fetch.

Two honest caveats. First, this is a Patreon-backed experimental node from a single developer (the pack is literally his "development takeaways" - snippets extracted from his ETUR upscaler), so the docs are the tooltips and little else. Second, detail_enhancer at non-zero can visibly slow a 20+ step run, so start at zero, tune the inpainting window first, and only then dial in the fancy controls. For plain inpainting with a defined injection window, this is a genuinely nicer drop-in than stitching together the stock nodes.