Research map / Image generation / Editing and style transfer
Face2QR: A Unified Framework for Aesthetic, Face-Preserving, and Scannable QR Code Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Face2QR is a training-free three-stage pipeline that generates personalized QR codes preserving face identity, aesthetics, and scannability using SDXL with InstantID and QR control networks, face-aware module reshuffling, and latent optimization with spatially dynamic loss.
Problem
Generating aesthetic QR codes that embed a human face while keeping the face identity recognizable and the code scannable, which existing image/style transfer and generative methods trade off poorly.
Approach
Combines SDXL with InstantID (face ID) and a QR controller in a training-free pipeline; reshuffles QR modules inside the face region (keeping face-region modules fixed and re-encoding error correction) to resolve face/QR conflicts; and enhances scannability by optimizing the VQ-VAE latent with a spatially dynamic code loss that is gentler on the face region plus a Wasserstein aesthetic content los
Key results
- Face2QR achieves the highest ArcFace face feature distance d=0.51 versus 0.50 (ArtCoder) and 0.43 (Text2QR), and AesBench score 90.1 versus 87.5 (Text2QR) and 62.0 (ArtCoder).
- Scan success rate averages above 94% across three sizes, two angles and three scanners.
- IDRS reduces error rate of the first-stage image by more than half; the initial output has over 43% error.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Face2QR instead of ArtCoder | Face2QR · Face images and generated QR images | face feature distance: 0.50 -> 0.51 | Helped |
| Use Face2QR instead of Text2QR | Face2QR · Face images and generated QR images | face feature distance: 0.43 -> 0.51 | Helped |
| Use Face2QR instead of ArtCoder | Face2QR · Face images and generated QR images | Aesbench score: 62.0 -> 90.1 | Helped |
| Use Face2QR instead of Text2QR | Face2QR · Face images and generated QR images | Aesbench score: 87.5 -> 90.1 | Helped |
| Use Face2QR instead of Text2QR | Face2QR · QR scanning across tested decoders, sizes, and angles | success rate | Mixed |
| Add ID-aware QR ReShuffle and stage 2 generation (ablation) | Face2QR · QR image facial region | error modules: over 43% | Helped |
| Add ID-aware QR ReShuffle instead of guiding stage 2 with a normal QR code (ablation) | Face2QR · QR image facial region | error modules | Helped |
| Use spatially adaptive Gaussian variance instead of uniform variance (ablation) | Face2QR IDSE · QR image facial region | face distortion | Helped |
| Use spatially adaptive loss weights instead of a uniform loss weight (ablation) | Face2QR IDSE · QR image | error modules | Helped |
| Use adaptive code loss with smaller face-region weight and variance and larger background-region weight and variance (ablation) | Face2QR IDSE · QR image face identity and scannability | face-region modification | Helped |
Limitations
- Constrained by generative model failures and inconsistent outputs, including conflicts between face image and prompt.
- Requires substantial computing power for high-resolution generation.
- Small evaluation set (200 comparison images, 20 codes for scanning, 30-user study) and no error bars.
Experimental setup
Datasets: MS-COCO (VGG-19 pre-training)
Benchmarks: AesBench
Metrics: face feature distance (ArcFace cosine similarity) AesBench aesthetic score scan success rate error module count user preference
Baselines: QArt Halftone QR ArtCoder Text2QR InstantID
Builds on: Stable Diffusion XL InstantID QR Controller ControlNet
Closest papers by meaning
- Text2QR: Harmonizing Aesthetic Customization and Scanning Robustness for Text-Guided QR Code Generation (CVPR 2024 · similarity 0.89)
- $\text{ID}^3$: Identity-Preserving-yet-Diversified Diffusion Models for Synthetic Face Recognition (NeurIPS 2024 · similarity 0.83)
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.82)
- FaceA-Net: Facial Attribute-Driven ID Preserving Image Generation Network (AAAI 2025 · similarity 0.82)
- Data Synthesis with Diverse Styles for Face Recognition via 3DMM-Guided Diffusion (CVPR 2025 · similarity 0.82)
- PortraitBooth: A Versatile Portrait Model for Fast Identity-preserved Personalization (CVPR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.