RegionalSamplerAdvanced

Regional sampling is how you tell one image "put this on the left and that on the right" and have it actually stick. You define a mask per region and a prompt (and sampler) per region, and the sampler paints each area with its own conditioning. RegionalSamplerAdvanced is the step-controlled version of Impact Pack's RegionalSampler - same idea, but you drive it with explicit start/end steps instead of a single denoise value, which is what you want when you're layering it into a bigger multi-stage sampling schedule.

The thing that makes Impact Pack's regional sampling distinct from just masking two generations together: per the README, "unlike TwoSamplersForMask, sampling for each region is applied during each step." The regions co-develop rather than one finishing before the other starts, so they blend at the boundaries instead of colliding.

How it works

You build REGIONAL_PROMPTS upstream (each is a mask plus a sampler, made with the RegionalPrompt node), hand them to this node along with a base sampler and a starting latent, and it runs the schedule: at each step the base sampler samples the whole frame, then each region's bound sampler samples its masked area. Two controls keep it clean - overlap_factor blends each region into the area outside its mask, and restore_latent, per the README, restores "the areas outside the mask to the base latent, preventing additional noise from being introduced outside the mask during region sampling."

The inputs that matter

• base_sampler (a KSAMPLER_ADVANCED) and regional_prompts - the base pass and the per-region prompt/sampler bundles. The core wiring.
• latent_image - what you're sampling.
• steps, start_at_step, end_at_step - the step schedule, which is the whole reason to use the "Advanced" variant over the plain one.
• overlap_factor (default 10) - region-to-surroundings blend.
• restore_latent (default true) - keep outside-mask areas pinned to the base; leave it on unless you have a specific reason.
• add_noise, noise_seed, return_with_leftover_noise - noise control for chaining passes.

The additional_mode / additional_sampler / additional_sigma_ratio inputs exist for a specific reason the README spells out: the sde and uni_pc samplers inject extra noise each step, so the node applies an additional corrective sampler (dpmpp_fast or dpmpp_2m) to keep regions stable. Leave them at defaults unless you're chasing that interaction. Output is a LATENT - VAE-decode it as usual.

How to install it

ComfyUI Manager: search ComfyUI Impact Pack, install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack, install requirements into ComfyUI's Python (pip install -r requirements.txt, or ..\..\..\python_embeded\python.exe -m pip install -r requirements.txt on Windows portable), restart. No auto-install since v7.6. A small SAM model downloads to ComfyUI/models/sams on first run. Regional prompts also pair well with the Inspire Pack, so many regional workflows want it installed too.

Common issues

A cryptic comfy.samplers error on run. This is the recurring one users hit with the RegionalSampler family - errors like module 'comfy.samplers' has no attribute 'calculate_sigmas_scheduler' mean your Impact Pack and ComfyUI core versions are out of sync. This pack pins hard to specific ComfyUI versions (the README is wall-to-wall version notices). Update both to current and it clears; downgrading one to match the other also works in a pinch.

Parameter order changed under you. The README flags that a past RegionalSampler update reordered parameters and broke old graphs. If you loaded an old workflow and it misbehaves, re-check the inputs rather than assuming the node is broken.

Regions bleed into each other. Tune overlap_factor down for crisper boundaries, and keep restore_latent on so region sampling doesn't leak noise outside its mask.