Research map / Image generation / Editing and style transfer
Learning Input-agnostic Manipulation Directions in StyleGAN with Text Guidance
Yoonjeon Kim, Hyunsu Kim, Junho Kim, Yunjey Choi, Eunho Yang
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| Learn a dictionary using multi-channel manipulation effects | StyleGAN2-ADA text-guided manipulation · FFHQ, 10–100 channels | R_CLIP: 0.224 -> 0.242; 0.228 -> 0.249; 0.232 -> 0.251; 0.237 -> 0.254 | Improved |
| Learn the dictionary using multi-channel manipulation effects | StyleGAN2-ADA direction recovery · FFHQ, 100 channels | MSE direction reconstruction: 0.0315 -> 0.0057 | Improved |
| Learn the dictionary using multi-channel manipulation effects | StyleGAN2-ADA direction recovery · FFHQ, zero direction channels | MSE disentanglement: 0.006218 -> 0.00486 | Improved |
| Use Multi2One instead of StyleCLIP GlobalDirection | StyleGAN2-ADA text-guided manipulation · FFHQ, 57 text prompts | R_CLIP: 0.2482 -> 0.2617 | Improved |
| Use Multi2One instead of StyleCLIP Local Optimization | StyleGAN2-ADA text-guided manipulation · FFHQ, 57 text prompts | R_CLIP: 0.2698 -> 0.2617 | Hurt |
| Add the multi-channel CLIP reconstruction loss (λ=0.01) (ablation) | Multi2One · FFHQ, 10–100 channels | R_CLIP: 0.235 -> 0.242; 0.240 -> 0.249; 0.243 -> 0.251; 0.248 -> 0.254 | Improved |
| Use supervised InterFaceGAN directions instead of unsupervised directions (ablation) | Multi2One dictionary learning · FFHQ, 100 channels | R_CLIP: 0.254 -> 0.228 | Hurt |
| Use top-k sparsity levels 10, 30, and 50 rather than including 100 (ablation) | Multi2One dictionary learning · FFHQ, 57 text prompts | R_CLIP increase: 0.121 -> 0.141 | Improved |
| Use top-k levels 10 and 30 only rather than 10, 30, and 50 (ablation) | Multi2One dictionary learning · FFHQ, 57 text prompts | R_CLIP increase: 0.141 -> 0.140 | No clear effect |
| Use top-k levels 10, 30, and 50 rather than top-k 10 only (ablation) | Multi2One dictionary learning · FFHQ, 57 text prompts | R_CLIP increase: 0.106 -> 0.141 (λ=0.01); 0.125 -> 0.115 (λ=0.1) | Mixed |
Closest papers by meaning
- StyleCLIP: Text-Driven Manipulation of StyleGAN Imagery (ICCV 2021 · similarity 0.92)
- Text-Guided Human Image Manipulation via Image-Text Shared Space (TPAMI 2021 · similarity 0.89)
- Unleashing In-context Learning of Autoregressive Models for Few-shot Image Manipulation (CVPR 2025 · similarity 0.86)
- Learning to Manipulate Artistic Images (AAAI 2024 · similarity 0.86)
- TextCtrl: Diffusion-based Scene Text Editing with Prior Guidance Control (NeurIPS 2024 · similarity 0.86)
- Prompt Augmentation for Self-supervised Text-guided Image Manipulation (CVPR 2024 · 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.
