Triverse AI has officially released its portrait-retouching diffusion model, MirrorPPR.
MirrorPPR introduces a new paradigm of Exemplar-Based Portrait Photo Retouching: instead of relying on vague text instructions, the model automatically infers retouching operations — facial-feature adjustment, face-shape refinement, body-proportion reshaping, and more — from a single before/after exemplar pair, then transfers them to new portrait images while preserving identity and non-edited regions.
Key Innovations
- Retouching Operation Extractor: a frozen MAE plus a trainable R-Former that distills the subtle difference signal between exemplar pairs
- Self-augmentation paradigm: the same spatial augmentation is applied to exemplar and query pairs, keeping operations strictly aligned
- MirrorPPR47M dataset: over 47 million portrait-retouching sample pairs covering 27 professional retouching operations
Experimental Results
On the SimFace-100 and ProPortrait-500 cross-identity benchmarks, MirrorPPR outperforms leading multi-reference / exemplar-guided / text-guided editing methods — including Qwen-Image-Edit, FLUX.2-dev, Nano Banana 2, and Seedream 4.5 — across PSNR, SSIM, LPIPS, and Face Similarity, reaching a Face Similarity of 0.96, far ahead of the runner-up (0.81).
MirrorPPR is now being integrated into our image-generation product line and will roll out to enterprise customers.