K2 Regional Face Detail

K2 Regional Face Detail is the node that keeps multi-character LoRA routing honest. In a full-body composition, each face covers only a handful of latent tokens - Krea 2's image tokens are 16×16 output pixels, so a face in a full-body frame is a rounding error. Even with perfect regional routing, the identity gets lost because there simply aren't enough tokens to carry it. This node fixes that the old-school way: it detects faces, assigns each one to the subject region its box sits in, and re-samples a padded crop at full resolution using only that region's LoRAs, then composites it back with a feathered mask.

It's the face-refinement stage of the K2 Lab suite in comfyui-FVMtools, and it's the difference between "two characters, kind of" and "two recognizable characters." For character-consistency work this is the node you'll actually reach for every run.

The inputs that matter

• image - the decoded render to inspect and refine.
• model - the base Krea model, not the K2 Compose output. The tooltip is explicit: regional gating makes no sense on an isolated crop, so connect the original model.
• clip, vae, plan - the encoder, the VAE, and the plan from K2 Compose (which supplies regions, priorities, prompts, and LoRA assignments).
• denoise - low values keep facial structure, higher values let the LoRA reshape the face. Start around 0.35.
• crop_size - working resolution per face crop (default 512).
• padding - expands the detected face box before cropping, so you get hair and context (default 2).
• feather and blend - the crop's soft border (as a fraction of crop size) and opacity. blend 0 keeps the original.
• lora_scale - an extra multiplier on the region LoRAs during the crop pass. This is the "turn the character's identity up for the face" knob.
• detector - yolo (any Ultralytics face model, recommended) or nanodet_onnx (the face_det.onnx used by K2Lab). Both are typically already in an install - no extra dependency. detector_model picks the Ultralytics file; threshold rejects background faces.
• cfg, sampler_name, scheduler - guidance and schedule for each crop pass. 1.0 CFG for Turbo, euler/simple.

Optional: detector_path (explicit path to face_det.onnx; empty uses FantasyPortrait auto-discovery) and require_region_lora (only refine regions that actually have a regional LoRA assigned).

Outputs: image (refined), face_mask, and a report of what was detected and refined. Each face gets a deterministic seed offset so different faces don't share identical noise.

The gotcha

Connecting the Compose-patched model here is the classic mistake - it renders crops that look wrong and you'll blame the LoRA. Use the original model. Everything else is tuning: if identity still washes out, raise lora_scale or denoise, don't chase it with the base pass.

Install

Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/ping1979ping/comfyui-FVMtools

Restart ComfyUI, find it under FVM Tools/K2. No extra Python dependencies, and the face detector is whatever your workflow already has.