Research map / Image generation / Autoregressive and latent generation
Domain Expansion of Image Generators
Yotam Nitzan, Michaël Gharbi, Richard Zhang, Taesung Park, Jun-Yan Zhu, Daniel Cohen-Or, Eli Shechtman
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Expand the generator with StyleGAN-NADA instead of using a separate adapted generator | StyleGAN2 · StyleGAN-NADA image comparison | User preference: 41.2% -> 58.8% | Improved |
| Expand the generator with MyStyle instead of using a separate adapted generator | StyleGAN2 · MyStyle image comparison | ID: 0.80 ± 0.06 -> 0.76 ± 0.05 | No clear effect |
| Expand the generator with MyStyle instead of using a separate adapted generator | StyleGAN2 · MyStyle image comparison | Diversity: 3.08 ± 0.15 -> 3.14 ± 0.14 | No clear effect |
| Expand the generator with StyleGAN-NADA instead of using a separate adapted generator | StyleGAN2 · StyleGAN-NADA image comparison | Diversity: 2.42 ± 0.13 -> 2.42 ± 0.13 | No clear effect |
| Expand the source generator with new domains while preserving its base subspace | StyleGAN2 · FFHQ | FID: 2.77 -> 2.80 | No clear effect |
| Expand the source generator with new domains while preserving its base subspace | StyleGAN2 · AFHQ Dog | FID: 7.43 -> 7.51 | No clear effect |
| Expand the source generator with new domains while preserving its base subspace | StyleGAN2 · LSUN Church | FID: 3.92 -> 3.76 | No clear effect |
| Expand the source generator with new domains while preserving its base subspace | StyleGAN2 · SD-Elephant | FID: 2.30 -> 2.70 | No clear effect |
| Increase the expansion from 105 to 400 new domains (ablation) | FFHQ StyleGAN2 · FFHQ base subspace | FID: 2.80 -> 2.83 | No clear effect |
| Use a larger distance between base and repurposed subspaces (ablation) | FFHQ StyleGAN2 expansion · New-domain CLIP alignment | CLIP error: s=5; s=20; s=50 | Improved |
| Set the base-to-repurposed distance too large (ablation) | FFHQ StyleGAN2 expansion · New-domain generation | Image quality: s = 50 | Hurt |
| Set the base-to-repurposed distance too small (ablation) | FFHQ StyleGAN2 expansion · New-domain generation | Target-domain representation: s ≤ 5 | Hurt |
Closest papers by meaning
- Image is All You Need to Empower Large-scale Diffusion Models for In-Domain Generation (CVPR 2025 · similarity 0.86)
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.86)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.86)
- Return of Unconditional Generation: A Self-supervised Representation Generation Method (NeurIPS 2024 · similarity 0.86)
- Understand Before You Generate: Self-Guided Training for Autoregressive Image Generation (NeurIPS 2025 · similarity 0.85)
- Elucidating the design space of language models for image generation (ICML 2025 · similarity 0.85)
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.
