Detail SEGS as Regions

You know the classic detailer loop: detect a region, crop it out, resample the crop at proper resolution, paste it back. It works, but it samples each region in isolation - and on a big canvas, isolated crops can drift from the overall scene. Detail SEGS as Regions deliberately takes the other road. It keeps the full image in one pass, uses global conditioning across the whole canvas, and pairs each SEG with its own ordered regional conditioning. Instead of N separate crop-refine cycles, you get one coherent pass where every region gets personal prompting but still answers to the whole image.

It's the pack's answer to a specific complaint about crop-based detailers: they fix faces, but sometimes they fix them out of context. If you've ever seen a detail pass where the improved hand no longer matches the lighting of the body it's attached to, that's the isolated-crop failure. Detail SEGS as Regions is designed so regions stay aware of the full scene.

How it works

The node takes the image, the segs regions, a model and vae, plus global positive and negative conditioning and a region_positive CONDITIONING_BATCH matched to the incoming SEGS order. The key control is global_prompt_weight (default 0.25): inside covered regions, this much of the prediction comes from the global positive prompt, and the rest comes from the regional prompt. Dial it up when you want regions to stay glued to the global scene; dial it down when a region needs to go its own way.

The rest is standard detail-pass plumbing: scale_factor (default 1, up to 5) enlarges each region's contribution for more detail room, upscale_method defaults to lanczos, and the usual sampler controls - seed, steps (20), cfg (8), sampler_name, scheduler, denoise (0.5). Compositing is governed by feather (5) for mask-edge blending, plus noise_mask (on) and noise_mask_feather (20) so sampling noise stays inside the region and unchanged pixels stay stable. tiled_encode / tiled_decode (both off) tile the VAE for big crops at a speed cost. Output is one image with the detailed regions blended back in.

The comparison you actually need

Crop-based (Detail SEGS by Scale Factor) samples each region at its own scale in isolation - great for fixing a small face in an otherwise-fine image, and the closer cousin to ADetailer/FaceDetailer. This node trades that isolation for coherence: one MultiDiffusion-style pass, global context preserved, regional conditioning per SEG. The tradeoff is that it's heavier (it's sampling the full canvas, not just crops) and less forgiving on region masks. Choose by goal: quick surgical fixes → scale-factor detailer; coherent scene-wide refinement with per-region prompts → this one.

Install

Ships in the SimpleSyrup pack. ComfyUI Manager → search SimpleSyrup → Install → restart, or manually:

Set-Location ComfyUI\custom_nodes
git clone https://github.com/Artificial-Sweetener/SimpleSyrup.git
Set-Location SimpleSyrup
..\..\venv\Scripts\python.exe -m pip install -r requirements.txt

Restart ComfyUI; the pack needs a current build (v3 extension API).

Common issues

• Region prompts pair with the wrong region - region_positive must match SEGS order exactly.
• Detail pass looks too strong/weak - that's denoise, then global_prompt_weight.
• Seams at region edges - raise feather and noise_mask_feather.
• Slow - it's sampling the whole canvas; enable tiled_encode/tiled_decode for large images.

It's the detailer for when per-region prompting matters more than per-region cropping.