AIRI shows KandiShop method: face similarity in SI images rises from 0.519 to 0.748
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
KandiShop fixes faces and packaging in finished images without training a model and in about 5 seconds on a single H100.
AIRI's controllable generative SI and video generation teams have published the KandiShop method. It fixes a face or product packaging in an already generated image without training, inversion or additional encoders. Inside the mask, two streams are computed — scene and identity; they are blended by the mask, and adaptive instance normalization (AdaIN) matches the identity's color to the scene. The reference is aligned to the frame using 5 keypoints for faces and, for products, SIFT matches together with Grounding DINO and SAM. According to the authors' measurements, on FLUX.2 klein 9B face similarity by ArcFace rose from 0.519 to 0.748 — above 0.571 for Nano Banana Pro. OCR F1 on packaging text was 0.841 versus 0.819 for GPT Image 2 High on 32 products from Products-10K. Processing one face or product takes about 5 seconds on an H100. The code is open for the FLUX.2 klein 9B version; the method was also tested on FLUX.2-dev, Qwen-Image-Edit-2511 and HunyuanImage 3.0. The authors position it as post-processing for third-party models, including generations from Nano Banana Pro.
Sources 1