Multi-character identity binding in stylized FLUX workflows: leakage, wrong-face refines, speed
Dry_Preparation_6762 · reddit · 2026-10-17
A developer describes their ComfyUI pipeline for rendering 1-4 characters in a consistent stylized look (custom style LoRA), each bound to a real reference photo. Single-character scenes work; multi-character scenes break:
Problems
- Identity leakage: global reference conditioning bleeds hairstyle/accessory features across characters and background figures;
- Wrong face in the refine pass: detection + embedding similarity is unreliable on stylized faces, putting character A's face on B;
- Head-and-shoulders references push toward cropped bodies or missing limbs;
- Speed target: 15-20s per final image including refinement on a single 24-32GB card.
Candidates under consideration: pre-defining each character's box and binding identity to the region during generation; PuLID-Flux with per-character attnmask on FLUX.1-dev; FLUX 3 Image Klein's bounding boxes with references.
Questions asked: what actually works for region-bound identity with 2-4 real people; Klein vs PuLID likeness; Crop & Stitch vs mask-filtered SEGS for region-local refine; permissively-licensed face detectors that work on stylized characters; acceleration tips (Nunchaku, TeaCache/MagCache, SageAttention, NVFP4, batching).
More from Multimodal
- AI video demo 'Omni 1'1 flash' wows Reddit with near-photoreal cinematic footage — Dense-Instruction210 · 2026-10-17
- Redditor recreates Austin Powers scene with AI video, 'Mini Me' likeness impresses — Alternative_Can_8663 · 2026-10-17
- N8Vision: open-source image browser for managing 50k+ AI-generated images — Consistent_Help2177 · 2026-10-17
- Dev trains a 'microdrama model': 'my wallet is gonna thank me' — Kyrannio · 2026-10-17
- Jeffrey Emanuel ships agent skill turning audio into spectrogram karaoke videos — doodlestein · 2026-10-17
- Video Rebirth ships BACH 1.0 Pro, an AI video engine for cross-shot character consistency — lmoroney · 2026-10-17