[RELEASE] ComfyUI-PuLID-Flux2 — First PuLID for FLUX.2 Klein (4B/9B)
First custom node enabling PuLID face consistency for the new FLUX.2 Klein architecture, rebuilt from scratch.
Developer iFayens has released ComfyUI-PuLID-Flux2, the first custom node to bring PuLID (Prompt-driven Latent Identity) face consistency technology to Stability AI's FLUX.2 Klein models. This release solves a critical problem in AI image generation: maintaining consistent facial identity across different scenes and prompts. Previously, using the same prompt with FLUX.2 Klein would generate different people in each image, but with this new node, users can now generate the same person consistently across diverse settings like Japanese streets, deserts, New York, Alps, and Bali.
What makes this release significant is that existing PuLID nodes only supported Flux.1 Dev, requiring a complete architectural rebuild for FLUX.2 Klein. The new architecture features different block structures (5 double/20 single blocks vs 19/38 in Flux.1), shared modulation instead of per-block, hidden dimensions of 3072 (vs 4096), and uses Qwen3 text encoder instead of T5. The node is currently functional using Flux.1 PuLID weights with partial compatibility for Klein 9B, though quality is slightly lower than without PuLID. The repository includes training scripts for developing native Klein-trained weights, and the developer welcomes contributions to improve the model's performance.
- First PuLID implementation for FLUX.2 Klein (4B/9B) enabling consistent facial identity across diverse image generations
- Completely rebuilt architecture required due to FLUX.2's different block structure (5 double/20 single blocks) and Qwen3 text encoder
- Includes training scripts for native Klein weights and currently uses Flux.1 weights with partial compatibility
Why It Matters
Enables professional creators to maintain character consistency across marketing campaigns, storyboarding, and brand imagery using cutting-edge FLUX.2 models.