Research map / Image generation / Editing and style transfer
PCA-Based Knowledge Distillation Towards Lightweight and Content-Style Balanced Photorealistic Style Transfer Models
Tai-Yin Chiu, Danna Gurari
Check 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 |
|---|---|---|---|
| Use PCA-derived channel lengths instead of CKD's empirically chosen channel lengths (ablation) | PCA-distilled VGG model · Photorealistic style transfer | model parameters: Ours-VGG uses 64.3% parameters of Ours-VGG-CKD | Improved |
| Use global rather than image-dependent local eigenbases (ablation) | PCA-distilled models · PST | stylization strength | Improved |
| Set the first VGG student layer to 10 channels rather than 5 (ablation) | PCA-distilled VGG model · Photorealistic style transfer | distillation feasibility | Improved |
| Add high-frequency residuals to PhotoWCT (ablation) | PhotoWCT · Photorealistic style transfer | content loss: 17.7% | Improved |
| Add high-frequency residuals to the original CKD model (ablation) | CKD-original · Photorealistic style transfer | content loss: 25.2% | Improved |
| Use PCA-derived channel lengths with CKD instead of CKD's empirical lengths (ablation) | CKD · Photorealistic style transfer | image quality | No clear effect |
| Use 75% rather than 85% variance preservation for channel selection (ablation) | PCA-distilled VGG model · Photorealistic style transfer | content loss: 2.25e$^6$ -> 2.20e$^6$ | Improved |
| Use 75% rather than 85% variance preservation for channel selection (ablation) | PCA-distilled VGG model · Photorealistic style transfer | style loss: 4.77e$^4$ -> 4.90e$^4$ | Hurt |
Closest papers by meaning
- Ultrafast Photorealistic Style Transfer via Neural Architecture Search (AAAI 2020 · similarity 0.87)
- Domain-Aware Universal Style Transfer (ICCV 2021 · similarity 0.87)
- Optical Flow Distillation: Towards Efficient and Stable Video Style Transfer (ECCV 2020 · similarity 0.87)
- ArtBank: Artistic Style Transfer with Pre-trained Diffusion Model and Implicit Style Prompt Bank (AAAI 2024 · similarity 0.86)
- Joint Bilateral Learning for Real-time Universal Photorealistic Style Transfer (ECCV 2020 · similarity 0.86)
- SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 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.
