Research map / Image generation / Editing and style transfer
TransferI2I: Transfer Learning for Image-to-Image Translation From Small Datasets
Yaxing Wang, Héctor Laria, Joost van de Weijer, Laura Lopez-Fuentes, Bogdan Raducanu
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 |
|---|---|---|---|
| add source-target initialization (finetune pretrained GAN on source and target) (ablation) | TransferI2I without self-init · Animal faces | mFID: 137.11 -> 114.23 | Improved |
| add self-initialization of adaptor layers (ablation) | TransferI2I without source-target init · Animal faces | mFID: 137.11 -> 122.12 | Improved |
| add self-initialization on top of source-target initialization (ablation) | TransferI2I (DeepI2I-like) · Animal faces | mFID: 114.23 -> 109.7 | Improved |
| add source-target initialization on top of self-initialization (ablation) | TransferI2I · Animal faces | mKID x100: 10.03 -> 9.40 | Improved |
| add auxiliary generator to source-target + self-initialization (ablation) | TransferI2I · Animal faces | mFID: 109.7 -> 103.55 | Improved |
| share more ResBlock layers between generator and auxiliary generator (1 -> 4) (ablation) | TransferI2I · Animal faces | mFID: 115.47 -> 103.55 | Improved |
| share more ResBlock layers between generator and auxiliary generator (1 -> 4) (ablation) | TransferI2I · Animal faces | mKID x100: 9.48 -> 9.25 | Improved |
| train TransferI2I architecture from scratch instead of transferring (ablation) | TransferI2I · Animal faces | mFID: 103.5 -> 356.1 | Hurt |
| TransferI2I versus DeepI2I on Animal faces | DeepI2I · Animal faces | mFID: 137.1 -> 103.5 | Improved |
| TransferI2I versus DeepI2I on Foods (larger dataset) | DeepI2I · Foods | mFID: 130.8 -> 107.8 | Improved |
| TransferI2I versus DeepI2I on cat2dog-200 dog to cat | DeepI2I · cat2dog-200 | FID: 83.71 -> 55.2 | Improved |
| TransferI2I versus StarGANv2 on larger cat2dog-2035 dog to cat | StarGANv2 · cat2dog-2035 | FID: 25.41 -> 27.0 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- DeepI2I — DeepI2I: Enabling Deep Hierarchical Image-to-Image Translation by Transferring from GANs (NeurIPS 2020) · better on 6 of 6 shared results
- StarGANv2 — StarGAN v2: Diverse Image Synthesis for Multiple Domains (CVPR 2020) · better on 3 of 4 shared results
- DRIT++ — DRIT++: Diverse Image-to-Image Translation via Disentangled Representations (IJCV 2020) · better on 3 of 3 shared results
- StarGAN — StarGAN: Unified Generative Adversarial Networks for Multi-Domain Image-to-Image Translation (CVPR 2018) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DeepI2I: Enabling Deep Hierarchical Image-to-Image Translation by Transferring from GANs (NeurIPS 2020 · similarity 0.96)
- Domain Adaptive Image-to-Image Translation (CVPR 2020 · similarity 0.91)
- ReMix: Towards Image-to-Image Translation With Limited Data (CVPR 2021 · similarity 0.91)
- TraVeLGAN: Image-To-Image Translation by Transformation Vector Learning (CVPR 2019 · similarity 0.91)
- Rethinking the Paradigm of Content Constraints in Unpaired Image-to-Image Translation (AAAI 2024 · similarity 0.90)
- Unsupervised Image-to-Image Translation Networks (NeurIPS 2017 · similarity 0.89)
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.
