Research map / Image generation / Editing and style transfer
G2L-CariGAN: Caricature Generation from Global Structure to Local Features
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.
Proposes G2L-CariGAN, a StyleGAN-based caricature generator that exaggerates global face shape on a 16x16 feature map and local facial features on a 32x32 feature map, plus a feature-map identity loss based on exaggerating the difference from the mean.
Problem
GAN-based caricature generation exaggerates only global facial structure in latent space, failing to highlight prominent local features (e.g., big eyes, hook nose) and often losing the subject's identity.
Approach
Replace vector latent codes for geometry with spatial feature maps (global 512x16x16 and local 512x32x32) inside a GLS latent space, exaggerate them with mask-based AdaIN per facial region, and use a cosine-similarity identity loss on feature maps relative to mean features (EDFM).
Key results
- Achieves FID 46.73 vs. 53.64 for CariME and 61.96 for WarpGAN (best among compared methods).
- Identity accuracy 97.65% on 294 photo-caricature pairs, highest in Table 2 (CariME 95.29%).
- Identity loss ablation: recognition rate rises from 66.83% to 72.55% with the caricature identity loss; user study identity recognition 74.59% for ours vs. lower for others.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the feature-map caricature identity loss (ablation) | G2L-CariGAN · Photo-to-caricature dataset | Recognition rate: 66.83 -> 72.55 | Helped |
| replace global feature map G with W+ latent codes (ablation) | G2L-CariGAN · Caricature generation | Hairstyle preservation | Hurt |
| replace local feature map L with W+ latent codes (ablation) | G2L-CariGAN · Caricature generation | Local facial feature exaggeration | Hurt |
| remove local exaggeration (ablation) | G2L-CariGAN · Caricature generation | Local feature emphasis | Hurt |
Surprising findings
- Off-the-shelf ArcFace achieves 99.46% on photo-to-photo but only 70.92% on photo-to-caricature, showing face recognition networks fail on caricatures.
Limitations
- Identity and FID evaluations rely on small test sets (294 pairs; 1800 photos) and a self-defined identity threshold K=0.82.
- Some baselines (CariGANs) lack public code and were compared using results from project websites.
Experimental setup
Datasets: WebCaricature CelebA CariGANs test set
Benchmarks: WebCaricature photo-to-caricature dataset
Metrics: FID identity accuracy (%) user perceptual preference
Baselines: CariGANs WarpGAN AutoToon Semantic-CariGANs CariME StyleCariGAN CariPainter DualStyleGAN
Builds on: StyleGAN pSp encoder DualStyleGAN StyleCariGAN AdaIN
Closest papers by meaning
- WarpGAN: Automatic Caricature Generation (CVPR 2019 · similarity 0.87)
- CaricatureBooth: Data-Free Interactive Caricature Generation in a Photo Booth (CVPR 2025 · similarity 0.85)
- CG-GAN: An Interactive Evolutionary GAN-Based Approach for Facial Composite Generation (AAAI 2020 · similarity 0.84)
- BlendGAN: Implicitly GAN Blending for Arbitrary Stylized Face Generation (NeurIPS 2021 · similarity 0.84)
- Interpreting the Latent Space of GANs for Semantic Face Editing (CVPR 2020 · similarity 0.83)
- CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis (NeurIPS 2025 · similarity 0.82)
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.