RegionalSampler

You want a red-haired woman on the left and a blue-robed wizard on the right, in one generation, without them bleeding into each other. That's what RegionalSampler is for. You hand it a base sampler for the whole canvas plus a set of masked regions, each with its own sampler and prompt, and it samples every region during each step rather than as a separate pass afterward. That "during each step" is the whole point - it's what makes the regions actually blend into a coherent image instead of looking pasted together.

It lives in ltdrdata's Impact Pack, the same detect-crop-refine toolbox behind FaceDetailer. Regional sampling is a smaller, nerdier corner of the pack than the detailer stuff, and it shows: this is one of the fiddlier nodes ltdrdata ships. Half the threads you'll find about it are people trying to get it to run at all. Go in expecting to spend some time wiring.

How it works

Sampling happens per step. The base sampler advances the entire latent one step, then each regional sampler re-samples its masked area toward its own prompt, then the next step runs, and so on. Because the regions are refined in lockstep with the base rather than bolted on at the end, the seams stay soft. This is the key difference from TwoSamplersForMask, which finishes the base area completely before touching the masked one.

The regions themselves come from RegionalPrompt nodes (mask + an advanced sampler), combined with CombineRegionalPrompts if you have several.

The inputs and outputs that matter

• base_sampler (KSAMPLER_ADVANCED) - the sampler for everything outside your masks. Build it from KSamplerAdvancedProvider.
• regional_prompts (REGIONAL_PROMPTS) - your masked regions, one per subject. This is the input that makes the node do its job.
• samples (LATENT) - the latent you're sampling, usually an empty latent at your target size.
• base_only_steps (default 2) - how many opening steps run the base sampler alone before the regions kick in. A couple of steps here lets the overall composition settle first.
• overlap_factor (default 10) - how far each region reaches past its mask edge to blend with the surroundings. Raise it if you're getting hard borders.
• restore_latent (default on) - after each region samples, it restores everything outside that region's mask back to the base latent, so region sampling can't dump stray noise into the rest of the frame. Leave it on unless you have a reason not to.

The single output is a LATENT - feed it into a VAE Decode like any other sampler result.

One thing the README calls out and it's worth heeding: the sde and uni_pc samplers inject extra noise each step, so when a region uses them the node quietly applies a corrective sampler underneath. Not something you set, just don't be surprised the internals are doing work.

How to install it

Easiest path is ComfyUI Manager: open the Manager, search ComfyUI Impact Pack, hit Install, restart. Done.

Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Impact-Pack
cd ComfyUI-Impact-Pack
pip install -r requirements.txt

Run that pip inside the same Python environment ComfyUI uses (for the portable build that's ..\..\..\python_embeded\python.exe -m pip), then restart. Automatic install was dropped back in v7.6, so the Manager or the manual steps are the only routes now. First launch pulls a small SAM model into ComfyUI/models/sams - that's for other parts of the pack, not this node, and it's harmless.

Common issues & troubleshooting

The recurring one is a plain crash on execution, historically an error about calculate_sigmas_scheduler not existing - a version-mismatch between Impact Pack and ComfyUI core. The fix that's worked for people is boring: update Impact Pack to the latest version (and update ComfyUI while you're there). Regional sampling leans hard on core's sampler internals, so a stale pack against a newer ComfyUI, or the reverse, is the usual culprit.

If two regions still bleed into each other, that's an overlap and mask problem, not a bug - tighten your masks and lower overlap_factor. And note the README's own warning that RegionalSampler's parameter order changed in a past update, so any ancient example workflow you import may wire the seeds wrong; rebuild it rather than trusting old JSON.