Achieve character consistency without LoRA using portable .char files
ashishsanu · reddit · 2026-08-15
A new method achieves character consistency in FLUX.2 without training LoRA, using a portable .char file to bundle character traits.
Mechanism:
Combines DINOv2 (robust visual features for the whole subject) and SFace (face recognition signature), with YuNet for face detection. This data is packed into a .char file containing manifest, refs, derived data, and scoring signatures.
Workflow:
- Build: Input photos to extract and pack features.
- Generation: Select the character from a dropdown. The file injects references via FLUX.2's native multi-reference channel, avoiding re-attaching images.
Advantages:
- Persistent character definition vs. one-time injection (PuLID/FaceID).
- Covers the whole subject, not just the face.
- Scores every generation against stored signatures to detect drift.
Limitations: Profiles and stylized renders drift more; body consistency is weak; struggles with popular celebrities due to model conflicts. Currently supports only FLUX.2 family.
More from Multimodal
- How meloform works: WASM pitch detection plus differentiable circuits learning 1,022 sine coefficients — henloitsjoyce · 2026-10-16
- Documentary filmmaker urges peers to watch 'You Can See Everything' — kyliebytes · 2026-10-16
- Redditor's first Cumfy/Seedance video nails a cinematic one-shot generation — Wave_N_Wind · 2026-10-16
- Kling 4 returns after 8 months: 10 hyper-realistic AI ads for under $5 — CurieuxExplorer · 2026-10-16
- Shanghai AI Lab unveils Intern Lumina U2, a diffusion LLM unifying image, video and 3D tasks — Shanghai-AI-Laboratory · 2026-10-16
- NAVER AI Lab proposes Speculative Forcing to speed up autoregressive video generation — naver-ai · 2026-10-16