Person Detailer

If you've ever tried to put two or three different characters in one image and watched the model blend them into one mush-faced average, you know the pain this node exists for. Person Detailer is the workhorse of the FVMtools pack: it takes an image, a set of detected faces, and - here's the part FaceDetailer from Impact Pack can't do - applies a different LoRA and prompt to each person. Person one gets your character LoRA and a "short brown hair" prompt; person two gets the other LoRA and "long black hair"; everyone else can be swept up by a generic slot. It's the multi-character problem from character-consistency land, solved with a bunch of per-slot inpaint passes instead of hoping regional prompting cooperates.

What it actually does

The mechanism is the old detect-crop-refine loop - the same idea Impact Pack's FaceDetailer popularized in ComfyUI - but re-engineered for multiple identities. Feed it images plus person_data from Person Selector Multi (the node that matched each face to a reference and generated its masks), and it iterates over your reference slots one at a time. For each enabled slot it crops the matched region, applies that slot's LoRA as a temporary clone on your base model, encodes the slot's prompt, and runs a masked inpaint pass at target_width × target_height - 800×1200 by default, which is the whole point: a face that's 80×80 in the full frame gets a proper resolution budget. Then it feathers the edges (mask_blend_pixels, default 32) and stitches it back. Slots process sequentially, each inpainting into the result of the previous one, so the image builds up progressively.

There's real engineering under the hood. LoRA files live in a 6-slot LRU cache so the same LoRA isn't re-read from disk for every face. And before the batch loop it pre-computes every unique (LoRA, strength, prompt) combination once - five people sharing one LoRA means one load and one conditioning encode, not five. If you're running Z-Image Turbo (Lumina2), it auto-detects it and converts diffusers-style LoRAs to the fused QKV format on the fly. That last bit is why the defaults look weird: steps defaults to 4 and denoise to 0.52, tuned for distilled turbo models. On a regular SDXL or Flux checkpoint, raise steps before you conclude it's broken.

The inputs that matter

• reference_1_enabled / reference_1_lora / reference_1_prompt - slot 1 is on by default; slots 2–5 start disabled. Each slot's prompt, if empty, falls back to the positive_base conditioning. This is where you wire per-character identity.
• denoise - 0.52 default. Lower keeps more original detail; too high and you're re-drawing the face, not refining it.
• generic_enabled + generic_prompt - the catch-all. It starts off; flip it on and, with generic_catch_unprocessed (on by default) doing the sweeping, it details any face no reference slot claimed, so nobody gets left looking mushy.

Outputs: images (the final stitched result), plus refined, refined_references, and refined_generic - grids of the inpainted crops, handy for checking what each slot actually did. Wire images straight to a SaveImage or into an upscaler.

Installing it

Grab the whole pack once - every node in this family ships together. Via ComfyUI Manager, search comfyui-fvmtools, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools
pip install insightface>=0.7.3 onnxruntime-gpu>=1.17.0 opencv-python>=4.8.0 numpy>=1.24.0

Restart ComfyUI. InsightFace's buffalo_l models auto-download on first use, but the BiSeNet segmentation model does not - drop parsing_bisenet.pth (53 MB, from xinntao/facexlib releases) into ComfyUI/models/gfpgan/ or face/head masks silently won't exist.

Where people get burned

The classic gotcha: an empty output. If you see a 64×64 black placeholder image come out, that's not a crash - it means no face matched any enabled slot, so nothing got detailed. Check your reference images, the threshold on Person Selector Multi, and whether the right slots are enabled. Also worth knowing: this node is LOUD - it loads a face detector, a face parser, potentially SAM, and runs a full inpaint per person. First run on a big batch can feel like it hung while models initialize. It's not hung, it's downloading buffalo_l.