NAI Multi-Character Detailer

Normal detailers fix faces by re-rendering each one with your local sampler. This one fixes them by calling the NovelAI API. Same detect-crop-inpaint-paste loop, except the "inpaint" half is a paid remote request - and the part that makes it interesting is that each detected region gets the original character prompt it belongs to, plucked out of your multi-character setup, instead of one shared prompt smeared across every face.

That matters specifically for multi-character NovelAI work. If you generate four characters and their faces come back mushy, a plain FaceDetailer has one blunt instrument: one prompt for all four crops, which is how you get a room of clones. Here, a local tagger looks at each crop and asks "is this A or B?", then sends that character's own prompt - the exact text you typed, franchise tags and all - to the inpainting request.

How it works

The pipeline is the standard detailing loop with the API swapped in for the sampler. A detector bundle from the NAI Detection Combine Loader finds regions, then the node repairs each mask locally: fill small notches with a morphological close, fill enclosed holes, dilate. That repair is deliberately conservative - geometry only, no model, and added pixels are accepted only inside a face-shaped band and under a size budget, so a large occluded area never gets blindly painted over. Then it crops and sends the crop out as a Focused Inpainting request with the matched character's positive and negative prompt.

Matching is where the optional tagger comes in. With matching_mode=wd14 and the node pack's local PyTorch WD tagger connected, each region is scored against your original character slots. The tagger's output never becomes prompt text - it only selects a slot. That's the whole point: the character's own original prompt goes out verbatim. Weak matches are retried with a tighter crop (match_crop_factor) and skipped entirely if they stay ambiguous, and there is no manual mapping input, so a skipped face keeps its original pixels rather than getting the wrong identity.

The inputs that actually matter

Everything is set on the node itself except detection:

• image - the finished render.
• detectors - the NAI_DETECTION_BUNDLE output from the combine loader. This socket only accepts that type, which is why you can't feed Impact Pack SEGS straight in.
• characterPrompts + tagger + matching_mode - the trio that turns a detailer into a multi-character detailer. Multiple characters in a connected nai_context automatically flip matching_mode to wd14 and warn if no tagger is plugged in.
• strength (0.55), seed (-1, so random per region unless you pin it), and max_regions (5 by default, up to 64). The node's own warning is worth repeating: quality falls off as region count climbs.
• limit_opus_free - on by default, and the reason half of your first runs will stop before spending anything.

detail_mode controls how many regions get processed at all - all (the default, ordered left to right then top to bottom) or first, which takes the highest-priority detection and stops. Also on the panel: dilation (8px of extra editing room), mask_feather (8px blend, separate from crop context), crop_factor, thresholds, and save_format for the autosaves.

Outputs are image, mask_visualization (your check on what it actually decided to touch) and matching_report, a JSON string listing each region's assigned slot, score, margin and whether it was edited or skipped. Wire the image into a Save Image node; read the report when a face comes back untouched.

Install

Via Manager, search ComfyUI_RS_NAI_API_Request. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/raspie10032/ComfyUI_RS_NAI_API_Request.git
python -m pip install -r ComfyUI_RS_NAI_API_Request/requirements.txt

You also need ComfyUI-Impact-Pack and ComfyUI-Impact-Subpack for the detector provider, plus a .env in the pack folder with NAI_ACCESS_TOKEN=... before any region can be corrected. For character matching, install the optional tagger requirements with the same Python that runs ComfyUI (python -m pip install -r requirements-tagger.txt) and let the WD tagger loader pull its weights - about 377 MB for the default WD ViT v3, stored under models/rs_wd_tagger.

Where people get burned

Executing this node performs correction - the old preview-only switch is gone, so a workflow that still has that widget checked is no longer protected by it. With limit_opus_free=True the node allows only validated Focused Inpainting crops (smaller than the source, ≤1,048,576 pixels, ≤28 steps) and blocks everything else before it reads your token. Set that to False deliberately. The upstream examples ship with protection on, so a workflow that appears to do nothing is usually doing exactly what it was told.

Second trap: the pack is small and Korean-community-rooted, and NovelAI-in-ComfyUI barely registers on the big subreddits, so the repo's own docs are a better help source than any thread you'll find. The pattern those docs encode - detail at native resolution, then upscale - still applies.