Detail Enhancer (Sigmas)

Detail Enhancer (Sigmas) is the schedule-rewriting cousin of the pack's "Detail Enhancer (Ultimate)". Where the Ultimate version wraps a sampler object and rewrites the schedule internally, this one sits out in the open: it takes a SIGMAS schedule in, modifies it, and hands a new schedule back. Slot it between your scheduler and the KSampler's sigmas input and it does the same job without you having to learn a wrapped-sampler workflow.

The pitch is simple and genuinely useful: your sampler already spends a fixed number of steps across the whole denoising range, but most of the detail worth having lives in the middle - the transition zone between coarse shape and fine texture. This node spends more of your existing steps there and inflates the noise values in that zone so the sampler is forced to work harder at texture. No extra steps, no second pass, no hires-fix upscale. Just a reshaped schedule.

How it works

It runs three passes on the sigma curve, all confined to the zone between start_percent and end_percent (defaults 0.1–0.8, the meaty middle):

• Time dilation ("relativity warp"): the timeline is resampled so time effectively moves slower inside the zone - enhance_factor 2.0 means twice as many of your N steps land there. This is a CDF-inversion resample, so the total step count never changes; you're redistributing, not adding.
• Texture boost: sigma values inside the zone get inflated (1.0 input ≈ a 15% lift, scaling up from there), which makes the sampler "denoise harder" and carve sharper texture.
• Monotonicity healing: boosting sigmas can create an upward bump in the curve that breaks Euler/DPM-family samplers, so the node flattens those bumps on the way out.

If you set enhance_factor to 1.0 and texture_boost to 0.0, it returns the schedule untouched - a nice passthrough when you want to disable enhancement without rewiring.

The inputs that matter

• sigmas - in from your scheduler (or anywhere a SIGMAS lives).
• enhance_factor (default 2, 1–10) - how much slower time runs in the zone. The tooltip calls it "time dilation in the selected sampling zone." Start at 2; you rarely need more than 3–4.
• texture_boost (default 0.5, 0–5) - sigma inflation for sharpness. The tooltip is honest: above 1.0 is strong. 0.5 is a gentle crispening; past 1.0 you're courting crunchy oversharpening.
• start_percent / end_percent (defaults 0.1 / 0.8) - the zone bounds. The edges are soft (a smoothstep fade), so you don't get a hard seam in the result.

The single output, sigmas, feeds straight into a KSampler's sigmas input in place of the raw scheduler output.

Install and the V3 requirement

Part of ComfyUI-Flow-Assistor, so it's a pack-wide install:

cd ComfyUI/custom_nodes
git clone https://github.com/Merserk/ComfyUI-Flow-Assistor.git

Restart ComfyUI (or use Manager → search "Flow-Assistor"). Important: the pack is V3-only - you need a current ComfyUI build with the comfy_api.latest node API and Python 3.10+. No pip dependencies, no model files, and unlike the captioning node in this pack there's nothing to download.

Gotchas

The big conceptual one: this node redistributes your steps, it doesn't add them. If your total step count is already low, time dilation just makes the middle denser at the expense of the ends - early denoising (composition) and late denoising (cleanup) can get starved. Bump your step count a little if you push enhance_factor high. And remember texture_boost above 1.0 is explicitly flagged as strong; if your image starts looking crunchy or plastic, drop it before you touch anything else. If the whole thing feels like it's doing nothing, check that the schedule actually covers the zone you set - a narrow late start_percent on a fast schedule leaves almost nothing to warp.