Coming up Tomorrow! Flux2Klein Identity transfer
New ComfyUI nodes let you lock character identity while changing poses and backgrounds with 0.3-0.5 guidance strength.
Developer Capitan01R has released a significant update to the Flux2Klein-Enhancer extension for ComfyUI, introducing two specialized nodes that solve a persistent challenge in AI image generation: maintaining character identity while allowing creative freedom. The Identity Guidance node operates during the sampling loop with adjustable strength (0.3-0.5 recommended) and timing controls (start_percent/end_percent), offering three modes: adaptive for prompt-driven changes, direct for complete locking, and channel_match for color/feature palette transfer only. This handles macro-level identity correction while leaving room for natural texture refinement.
The Identity Feature Transfer node works at a more granular level, steering features inside attention blocks with cumulative strength (0.15-0.25 recommended) across 24 possible blocks (0-23). It employs three distinct modes: cosine_pull for per-feature matching, topk_replace affecting only the most similar tokens (with top_k_percent control), and mean_transfer for overall character flavor. The breakthrough comes from using attention output steering rather than scaling k/v layers, creating a similarity-gated system that pulls similar features closer while leaving dissimilar elements (like new backgrounds or poses) untouched.
Together, these nodes represent a technical leap in controllable generation, allowing artists to lock character identity deep in the model's processing blocks while granting Flux2Klein complete freedom to generate diverse poses, compositions, and backgrounds according to prompts. The workflow includes pre-configured settings and documentation, making this advanced technique accessible to ComfyUI users working with character consistency across multiple generated images.
- Two new ComfyUI nodes: Identity Guidance (0.3-0.5 strength range) and Identity Feature Transfer (0.15-0.25 strength per block)
- Uses similarity-gated attention output steering instead of rigid k/v layer scaling for more natural results
- Enables character identity locking while allowing full creative freedom for poses, backgrounds, and prompt following
Why It Matters
Enables consistent character generation across multiple AI images while maintaining creative flexibility for professional artists and content creators.