Research map / Image generation / Editing and style transfer
HyperEditor: Achieving Both Authenticity and Cross-Domain Capability in Image Editing via Hypernetworks
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 HyperEditor, which uses hypernetworks to generate weight factors that reassign a pre-trained StyleGAN2 generator's weights, guided by directional CLIP loss, enabling authentic attribute editing and cross-domain style transfer.
Problem
Latent-code-based StyleGAN2 editing yields authentic edits but is confined to the generator's image domain, so cross-domain style transfer and authentic attribute editing are hard to achieve together; paired pre-/post-edit data is also scarce.
Approach
Edit images by generating per-layer weight scaling factors with hypernetworks (conditioned on image features and CLIP text-difference embeddings via a fusion modulation module) rather than manipulating latent codes, supervised by a directional CLIP loss, and an adaptive layer selector that picks which generator layers need weight factors.
Key results
- Achieves highest CLIP score (27.35) among compared methods and best PSNR (25.33), LPIPS (0.22), SSIM (0.85), and FID-0 (6.19) in the face editing table.
- Adaptive layer selector reduces hypernetwork parameters by about 80% (71.48M to 15.66M) while staying close to full model quality (CS 26.37).
- In a 20-participant user study, HyperEditor received 51.75% of accuracy votes, versus 15% for the next best (StyleGAN-NADA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HyperEditor instead of the strongest reported baseline on CS | StyleGAN2 image editing · CelebA-HQ, ten facial attributes | CS: 26.69 -> 27.35 | Helped |
| replace directional CLIP guidance with Global-CLIP guidance (ablation) | HyperEditor · CelebA-HQ, ten facial attributes | CS: 27.35 -> 22.96 | Hurt |
| add adaptive layer selector (ablation) | HyperEditor · CelebA-HQ, ten facial attributes | Nop: 71.48M -> 15.66M | Helped |
| add adaptive layer selector (ablation) | HyperEditor · CelebA-HQ, ten facial attributes | SSIM: 0.85 -> 0.87 | Mixed |
| increase the adaptive-selector threshold to λ_std≥0.8 (ablation) | HyperEditor with adaptive layer selector · facial attribute editing, including smile | editing effectiveness | Hurt |
| use HyperEditor instead of StyleGAN-NADA | StyleGAN2 image editing · CelebA-HQ, ten facial attributes | ACC: 15.00% -> 51.75% | Helped |
Surprising findings
- Global CLIP loss (instead of directional) causes large global distortion and blurriness, with FID-1 rising to 61.07 and CLIP score dropping to 22.96.
Limitations
- Full style transfer is incomplete: edits only partially transform images to the target domain, since weight scaling cannot fully reach optimal weights.
- Adaptive layer selector only suits single-attribute single-model editing; multi-attribute use requires all layers.
- Relies on a hyperparameter lambda_std whose choice trades off edit strength and preservation; high values make edits ineffective.
Experimental setup
Datasets: FFHQ CelebA-HQ LSUN AFHQ
Benchmarks: FFHQ CelebA-HQ LSUN AFHQ
Metrics: PSNR LPIPS SSIM identity similarity (IDS) FID CLIP score user preference
Baselines: StyleCLIP-LM StyleCLIP-GD StyleGAN-NADA DeltaEdit Ours(Global-CLIP)
Builds on: StyleGAN2 CLIP StyleCLIP e4e hypernetworks
Closest papers by meaning
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.87)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · similarity 0.87)
- FunEditor: Achieving Complex Image Edits via Function Aggregation with Diffusion Models (AAAI 2025 · similarity 0.86)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.86)
- FeedEdit: Text-Based Image Editing with Dynamic Feedback Regulation (CVPR 2025 · similarity 0.86)
- AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea (CVPR 2025 · similarity 0.86)
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.