Image & Video

FLUX.2 Klein Identity Feature Transfer V3 (Final)

New V3 commit system reduces feature mush and background bleed in identity transfer.

Deep Dive

FLUX.2 Klein has released Identity Feature Transfer V3, the final and cleaner version of its identity transfer node. The update is built around three presets—MIDUM_LOCK, HARD_LOCK, and SOFT_LOCK—that override manual settings, making it much easier to use without deep knowledge of blocks and hooks. The key innovation is a commit system: instead of constantly averaging the generation toward the entire reference image, V3 finds the best reference match for each generation token. If that match stays stable, it commits to it, then applies a lighter anchor force afterward. This means less feature mush, less random background bleed, and cleaner identity preservation.

For users who need finer control, a custom preset exposes parameters: double_schedule and single_schedule (block-level injection strengths), double_sim and single_sim (matching strictness), commit_margin (how obvious the best match must be), commit_confirm (how many repetitions before locking), commit_anchor (residual identity pull after lock), and mask_threshold (for optional subject masks that restrict which reference tokens are sampled). The node still supports an optional subject mask to limit identity extraction when the reference has multiple subjects. Presets are intended for most use cases: start with MIDUM_LOCK, then HARD_LOCK if identity drifts, or SOFT_LOCK if the reference overpowers the generation.

Key Points
  • Three new presets (MIDUM_LOCK, HARD_LOCK, SOFT_LOCK) simplify identity transfer without needing to understand every block.
  • Commit system finds the best reference match per token, locks it, then reduces pull force, cutting feature mush and background bleed.
  • Custom mode offers detailed controls: block schedules (double/single), similarity thresholds, commit margin/confirm/anchor, and optional subject mask.

Why It Matters

Makes AI identity transfer more reliable and accessible, reducing artistry guesswork for creators.