Research map / Image generation / Autoregressive and latent generation
Exact Fusion via Feature Distribution Matching for Few-shot Image 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.
Proposes F2DGAN, a fusion-based few-shot GAN that fuses features via VAE-modeled feature distributions and exact histogram matching, improving quality and diversity on Flowers, Animal Faces and VGGFaces.
Problem
Few-shot image generation suffers from a quality-diversity trade-off; existing fusion methods match features inaccurately (via similarity or attention), causing artifacts in texture and structure.
Approach
Replace point-wise feature matching with exact histogram (eCDF) matching of feature distributions, where class-level feature distributions are learned by a VAE-based variational feature learning module, and fuse features via sort-matching with stop-gradient for training.
Key results
- F2DGAN achieves FID 38.26 on Flowers with LPIPS 0.4325, versus 34.59 FID for LSO (best FID on Flowers), with the best LPIPS among fusion-based methods.
- On low-data classification with augmentation, F2DGAN reaches 84.02% on Flowers, 60.76% on Animal Faces and 77.92% on VGGFaces, the best of all compared generators.
- Ablation on Flowers shows removing FDMM raises FID from 38.26 to 58.48 and removing VFLM drops accuracy from 84.02% to 72.88%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use F2DGAN instead of the strongest baseline for FID | Few-shot image generation · Flowers | FID: 34.59 -> 38.26 | Hurt |
| Use F2DGAN instead of the strongest baseline for FID | Few-shot image generation · Animal Faces | FID: 23.67 -> 25.24 | Hurt |
| Use F2DGAN instead of the strongest baseline for FID | Few-shot image generation · VGGFaces | FID: 3.98 -> 4.25 | Hurt |
| Add FDMM (ablation) | F2DGAN · Flowers | FID: 58.48 -> 38.26 | Helped |
| Add the image-matching reconstruction loss (ablation) | FDMM · Flowers | FID: 46.42 -> 38.26 | Helped |
| Add VFLM (ablation) | F2DGAN · Flowers | FID: 44.96 -> 38.26 | Helped |
| Add the feature reconstruction loss (ablation) | VFLM · Flowers | FID: 40.87 -> 38.26 | Helped |
Surprising findings
- Transformation-based methods (AGE, LSO, HAE) achieve better FID or LPIPS than F2DGAN on some datasets, though F2DGAN gives better downstream classification gains.
Limitations
- Transformation-based baselines outperform it on some FID/LPIPS entries (e.g., LSO FID on Flowers and VGGFaces).
- Reported gains rely on reproducing baselines (LoFGAN, WaveGAN, HAE) under the authors' environment.
Experimental setup
Datasets: Flowers Animal Faces VGGFaces
Benchmarks: Flowers Animal Faces VGGFaces
Metrics: FID LPIPS top-1 accuracy
Baselines: FIGR DAWSON MatchingGAN F2GAN LoFGAN WaveGAN AGE LSO HAE
Builds on: LoFGAN EFDM ACGAN WaveGAN
Closest papers by meaning
- A Closer Look at Few-Shot Image Generation (CVPR 2022 · similarity 0.88)
- Generating a Fusion Image: One's Identity and Another's Shape (CVPR 2018 · similarity 0.86)
- Few-Shot Image Generation via Cross-Domain Correspondence (CVPR 2021 · similarity 0.85)
- Transformation GAN for Unsupervised Image Synthesis and Representation Learning (CVPR 2020 · similarity 0.85)
- WeditGAN: Few-Shot Image Generation via Latent Space Relocation (AAAI 2024 · similarity 0.84)
- Diverse Rare Sample Generation with Pretrained GANs (AAAI 2025 · similarity 0.84)
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.