Research map / Image generation / Image restoration
QC-StyleGAN - Quality Controllable Image Generation and Manipulation
Dat Viet Thanh Nguyen, Phong Tran The, Tan M. Dinh, Cuong Pham, Anh Tuan Tran
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 |
|---|---|---|---|
| Remove distillation loss, no freezing (ablation) | QC-StyleGAN · FFHQ | FID sharp: 3.65 -> 11.32 | Hurt |
| Remove distillation loss, with freezing (ablation) | QC-StyleGAN · FFHQ | FID sharp: 3.65 -> 33.45 | Hurt |
| Finetune only the last 1 synthesis block (M=1) instead of 2 (ablation) | QC-StyleGAN · FFHQ | FID sharp: 6.38 | Hurt |
| Stack P=5 conv layers in DegradBlock instead of 3 (ablation) | QC-StyleGAN · FFHQ | FID sharp: 4.02 | Hurt |
| Reduce quality code size to Dq=8 with L=64 (ablation) | QC-StyleGAN · FFHQ | FID sharp: 4.54 | Hurt |
| AdaIN-style DegradBlock instead of PCA-based (ablation) | QC-StyleGAN · FFHQ | FID sharp: 4.41 | Hurt |
| Inject q by concatenating with z, w, or DualStyleGAN structure (ablation) | QC-StyleGAN · FFHQ | content match of sharp/degraded | Hurt |
| Train degraded-image StyleGAN2-Ada separately vs QC-StyleGAN | StyleGAN2-Ada · FFHQ degraded generation | FID: 4.38 -> 3.23 | Improved |
| QC-StyleGAN versus SG2-Ada on sharp generation | StyleGAN2-Ada · FFHQ sharp | FID: 3.48 -> 3.65 | No clear effect |
| QC-StyleGAN versus SG2-Ada on LSUN-Church sharp | StyleGAN2-Ada · LSUN-Church sharp | FID: 3.86 -> 3.61 | Improved |
| Use QC-StyleGAN for multiple-degradation restoration | image restoration baselines · CelebA-HQ multi-deg | NIQE, LPIPS: 8.97 -> 5.64 NIQE (vs MPRNet) | Improved |
| QC-StyleGAN restoration versus specialized methods on noise LPIPS | DnCNN · CelebA-HQ noise | LPIPS: 0.080 vs 0.183 | Hurt |
Closest papers by meaning
- ReGANIE: Rectifying GAN Inversion Errors for Accurate Real Image Editing (AAAI 2023 · similarity 0.89)
- Analyzing and Improving the Image Quality of StyleGAN (CVPR 2020 · similarity 0.89)
- Robust Unsupervised StyleGAN Image Restoration (CVPR 2023 · similarity 0.89)
- High-Resolution GAN Inversion for Degraded Images in Large Diverse Datasets (AAAI 2023 · similarity 0.88)
- StyleGAN2 Distillation for Feed-forward Image Manipulation (ECCV 2020 · similarity 0.88)
- The Devil is in the Details: StyleFeatureEditor for Detail-Rich StyleGAN Inversion and High Quality Image Editing (CVPR 2024 · 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.
