KandiShop raises AI face similarity to the original from 52% to 75% without fine-tuning
A neural network puts a person from a photo into a new scene, but the face comes out "almost right" and the text on a jar turns into pseudo-letter mush. AIRI's KandiShop repairs the finished image without fine-tuning: it redraws the face or label from a reference and leaves everything else untouched. For FLUX.2 klein 9B, face similarity by ArcFace rose from 52% to 75%, and text reading accuracy from 76% to 84%. It is a ready-made post-processing step for any model.
AI-processed from Habr AI; edited by Hamidun News
You upload a photo of a person to a generator and ask it to place them in a snowy forest. The result is a beautiful shot, but your friends say: "It looks like you, but it isn't you." Products have the same problem: the logo and the colors of the can are right, but the small print on the label turns into squiggles.
The AIRI team has proposed KandiShop, a method that fixes exactly this after generation and without fine-tuning the model. The main metric is ArcFace, a face-recognition neural network that scores the similarity of two photos (100% means the same person). For FLUX.
2 klein 9B, that score rose from 52% to 75%. For developers of photo shoots, advertising and online stores, this is a way to make results recognizable and legible without changing the underlying generator.
The model can copy, but its creativity gets in the way
Modern in-context models (FLUX.2, Qwen-Image-Edit, HunyuanImage 3, and the commercial Nano Banana and GPT Image) take a photo and a text prompt and draw the result right away. Previously, the model had to be fine-tuned on a dozen photos of the person (DreamBooth, LoRA), or a separate face encoder had to be plugged in. Now a single photo is enough, but accuracy suffers: hair color and face shape are preserved, while the shape of the eyes and nose and subtle asymmetry drift.
The reason is that in a single pass the model solves two tasks that pull in different directions: inventing the pose, lighting and composition, and transferring every facial feature and every letter. When it has to choose, the second task gets sacrificed. The authors checked that the problem is not a lack of "memory": on the prompt "make an exact copy," face similarity by ArcFace was 96.5–97.1%, and text on packaging was recognized with an accuracy (OCR F1) of up to 92%. But in a new scene with several people, similarity drops to 24–60%. The conclusion: the two tasks need to be separated.
How KandiShop works
The method runs after generation. The input is a finished image and a reference photo; the output is the same image with a corrected face or label. The model's weights are not touched. Only the path by which it turns noise into an image is changed.
First, a large crop is cut out around the face or product: in a full frame, a face takes up only 5–10% of tokens, and the background drowns out the edit. Next, the reference is aligned to the crop: for faces, by five keypoints (eyes, nose, corners of the mouth); for products, with the Grounding DINO detector, SAM segmentation and SIFT keypoint matching. The crop is then denoised in two streams within a single transformer.
The first stream sees the original crop and holds the pose, background and lighting. The second sees the aligned reference and "knows" what the face or the letters look like. After each block, the streams are blended by the object mask, so subsequent blocks adapt the face to the scene.
To keep the face from arriving with the reference's lighting (for example, warm indoor light in a cold fog), the color and contrast of its features inside the mask are matched to the scene, using AdaIN. For products this step is turned off, because it smears small text. The main control is denoise, the share of the denoising path that is run anew: 0.5–0.75 for most cases, and 1.0 for a full replacement of the head or product. The finished crop is pasted back, and the rest of the frame stays pixel-for-pixel identical.
What the measurements showed
On 100 FFHQ faces in multi-person scenes, similarity for FLUX.2 klein 9B rose from 52% to 75%. The gain is larger for Qwen-Image-Edit, from 24% to 66%, and for HunyuanImage 3.0, from 49% to 84%. For comparison: Nano Banana Pro scores 57% and GPT Image 2 High scores 86%. The overall image quality score (HPSv3) barely changes, meaning the frame is not degraded.
For products, the team tested 32 packages in 320 scenes. For FLUX.2 klein 9B, text reading accuracy rose from 76% to 84%, and the character error rate (CER) fell from 65% to 51%. That beats GPT Image 2 High (82%) and Nano Banana Pro (78%).
The method is not tied to its own generator. Nano Banana Pro images were run through KandiShop on FLUX.2 klein 9B: face similarity rose from 57% to 76%, and text accuracy from 78% to 83%. In a blind comparison, seven VLM judges rated KandiShop's head or product replacement as more natural in 76% of cases versus a "native" replacement.
Caveats and practical takeaway
FFHQ faces may have been part of the models' training data, so the numbers may differ on new people. The method requires additional detectors and a second pass over each crop, and if the crop and reference align poorly, the result suffers. The open code so far covers only the FLUX.2 klein 9B version.
Practical takeaway: instead of fine-tuning a LoRA for each user, you can add a post-processing step that fixes faces and labels after any model, including commercial APIs.
- Nov 18, 2026Sciforium pushes DeepSeek-V4.1-Flash on AMD to 1 129 tokens per second
- Nov 18, 2026One Chinese Comma Switched DeepSeek to Chinese in 94% of Responses
- Nov 18, 2026Microsoft releases MAI Code 2 coding model: 98% of GPT-6.1 Sol's quality for 40% of the price
- Nov 18, 2026AWS releases SMART: spin up a WhatsApp agent on AgentCore with one command
Key Questions Answered
What are the key details on The model can copy, but its creativity gets in the way?
Modern in-context models (FLUX.2, Qwen-Image-Edit, HunyuanImage 3, and the commercial Nano Banana and GPT Image) take a photo and a text prompt and draw the result right away. Previously, the model had to be fine-tuned on a dozen photos of the person (DreamBooth, LoRA), or a separate face encoder had to be plugged in. Now a single photo is enough, but accuracy suffers: hair color and face shape are preserved, while the shape of the eyes and nose and subtle asymmetry drift.
How KandiShop works?
The method runs after generation. The input is a finished image and a reference photo; the output is the same image with a corrected face or label. The model's weights are not touched. Only the path by which it turns noise into an image is changed. First, a large crop is cut out around the face or product: in a full frame, a face takes up only 5–10% of tokens, and the background drowns out the edit.
What the measurements showed?
On 100 FFHQ faces in multi-person scenes, similarity for FLUX.2 klein 9B rose from 52% to 75%. The gain is larger for Qwen-Image-Edit, from 24% to 66%, and for HunyuanImage 3.0, from 49% to 84%. For comparison: Nano Banana Pro scores 57% and GPT Image 2 High scores 86%. The overall image quality score (HPSv3) barely changes, meaning the frame is not degraded. For products, the team tested 32 packages in 320 scenes.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.