Research map / Architectures and efficiency / Pruning and efficient tuning
Master: Meta Style Transformer for Controllable Zero-Shot and Few-Shot Artistic Style Transfer
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 Master zero-shot rather than StyTr2 | Master · Style transfer test set (220 pairs) | style loss: 1.50 ± 0.69 -> 0.81 ± 0.31 | Helped |
| Use Master few-shot rather than StyTr2 | Master · Style transfer test set (220 pairs) | content similarity loss: 0.48 ± 0.22 -> 0.38 ± 0.12 | Helped |
| Add learnable scaling and shifting instead of vanilla residual connections (ablation) | Master · Style transfer test set (220 pairs) | content similarity loss: 0.59 ± 0.26 -> 0.41 ± 0.13 | Helped |
| Add learnable scaling and shifting instead of vanilla residual connections (ablation) | Master · Style transfer test set (220 pairs) | content loss: 5.50 ± 0.63 -> 4.20 ± 0.68 | Helped |
| Add layer normalization to the Transformer style encoder (ablation) | Master · Style transfer test set (220 pairs) | style loss: 0.81 ± 0.31 -> 0.93 ± 0.33 | Hurt |
| Use 1 Transformer layer instead of 3 at zero-shot inference (ablation) | Master · Style transfer test set (220 pairs) | style loss: 0.81 ± 0.31 -> 0.92 ± 0.40 | Hurt |
| Increase inner optimization updates during meta training from 1 to 4 (ablation) | Master · Fast adaptation, 10th iteration | style loss: 3.389 -> 1.811 | Helped |
| Share Transformer layer parameters across the stacked layers | Master · 512×512 inference | number of parameters (M): 25.14 -> 10.75 | Helped |
| Update only the style encoder rather than the whole model during few-shot adaptation (ablation) | Master · Few-shot style transfer | training convergence | Helped |
| Use Reptile rather than MAML for meta training (ablation) | Master · Meta training | training time: 3 days for MAML v.s. 5 hours for Reptile | Helped |
| Use a CLIP image encoder rather than Swin for image style transfer (ablation) | Master · Image style transfer | local detail stylization | Hurt |
| Increase the number of stacked Transformer layers at inference (ablation) | Master · 512×512 inference | speed (sec. / image): 0.024 -> 0.030 -> 0.038 | Hurt |
Closest papers by meaning
- Puff-Net: Efficient Style Transfer with Pure Content and Style Feature Fusion Network (CVPR 2024 · similarity 0.89)
- OmniStyle: Filtering High Quality Style Transfer Data at Scale (CVPR 2025 · similarity 0.89)
- ArtAdapter: Text-to-Image Style Transfer using Multi-Level Style Encoder and Explicit Adaptation (CVPR 2024 · similarity 0.89)
- Z*: Zero-shot Style Transfer via Attention Reweighting (CVPR 2024 · similarity 0.88)
- SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.87)
- StyleMaster: Stylize Your Video with Artistic Generation and Translation (CVPR 2025 · similarity 0.87)
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.