Multi-character identity binding in stylized FLUX workflows: leakage, face swaps and fixes
Dry_Preparation_6762 · reddit · 2026-10-17
A ComfyUI developer details the pain points of binding multiple real-face identities in a stylized FLUX workflow (custom style LoRA) with 1–4 characters per scene: identity leakage under global reference conditioning, face-refine passes putting character A's face on B (embedding similarity unreliable on stylized faces), framing artifacts from head-and-shoulders refs, and a 15–20s/image speed target on a 24–32GB card.
Candidate approaches under consideration: pre-assigning per-character bounding boxes and binding identity to regions during generation, PuLID-Flux with per-character attnmask on FLUX.1-dev, and FLUX 3 Image Klein's refimage + bounding-box binding. The author asks the community about current best practices for region-bound identity, Klein vs PuLID likeness, region-only refinement (Crop & Stitch vs mask-filtered SEGS), permissively licensed face detectors that work on stylized characters, and speed tricks (Nunchaku, TeaCache/MagCache, SageAttention, NVFP4, batching).
More from coding & agent
- Inside Perplexity's SPACE repo: Apache-2.0 agent sandbox with 1.2k stars — abhishek__AI · 2026-10-17
- Perplexity open-sources SPACE, a Firecracker microVM sandbox for AI agent code — abhishek__AI · 2026-10-17
- Sol plays Asterix from screenshots, holding each move until its next reply arrives — wzenus · 2026-10-17
- GPT-6 Sol plays rhythm games in real time, picking keys and tap times from frames — wzenus · 2026-10-17
- Sol picks speed and lane targets frame-by-frame from onboard camera while racing — wzenus · 2026-10-17
- Sooool: latency-aware agent plays Touhou in real time, repo and guide open-sourced — wzenus · 2026-10-17