Research map / Continual learning / Online and replay-based
Piggyback GAN: Efficient Lifelong Learning for Image Conditioned Generation
Mengyao Zhai, Lei Chen, Jiawei He, Megha Nawhal, Frederick Tung, Greg Mori
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 unconstrained filters (lambda=1/4) to pure factorization (ablation) | Piggyback GAN · Paired Task 2 maps->aerial | FID: 303.90 -> 171.04 | Improved |
| add unconstrained filters (lambda=1/4) to pure factorization (ablation) | Piggyback GAN · Paired Task 2 maps->aerial | Acc: 99.82 -> 97.99 | No clear effect |
| add piggyback filters reusing previous filters, versus a model of one quarter size trained alone (ablation) | Piggyback GAN · Paired Task 2 maps->aerial | FID: 285.71 -> 171.04 | Improved |
| increase the unconstrained filter proportion lambda (ablation) | Piggyback GAN · Paired image-conditioned tasks | FID | Improved |
| change the initial task T1 (segmentations->facades instead of semantic labels->street photos) (ablation) | Piggyback GAN · Paired Task 2 maps->aerial | FID: 171.04 -> 174.67 | No clear effect |
| change the initial task T1 (edges->handbags instead of semantic labels->street photos) (ablation) | Piggyback GAN · Paired Task 2 maps->aerial | FID: 171.04 -> 166.20 | No clear effect |
| use Piggyback GAN instead of sequential fine-tuning | continual paired generation · Paired 4 tasks | FID: 259.76 -> 137.87 | Improved |
| use Piggyback GAN instead of binary-mask piggyback | continual paired generation · Paired 4 tasks | FID: 178.16 -> 137.87 | Improved |
| use Piggyback GAN instead of Lifelong GAN (quality degradation over tasks) | continual paired generation · Cityscapes FID over tasks | FID: 122.52 -> 160.28 -> 221.07 | Improved |
| use Piggyback GAN instead of binary-mask piggyback | continual unpaired translation · Unpaired 4 translation tasks | FID: 149.51 -> 113.89 | Improved |
| use Piggyback GAN instead of binary-mask piggyback | continual style transfer · Unpaired 2 style transfer tasks | FID: 118.26 -> 106.95 | Improved |
| use Piggyback GAN instead of sequential fine-tuning | continual style transfer · Unpaired 2 style transfer tasks | Acc: 50.00 -> 77.97 | Improved |
Later papers that compared with this method
2 later accepted papers report PiggybackGAN as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Few-Shot Incremental Learning for Label-to-Image Translation (CVPR 2022 · better on 4 of 4 shared results)
- Hyper-LifelongGAN: Scalable Lifelong Learning for Image Conditioned Generation (CVPR 2021 · better on 4 of 4 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
- Hyper-LifelongGAN: Scalable Lifelong Learning for Image Conditioned Generation (CVPR 2021 · similarity 0.92)
- Lifelong GAN: Continual Learning for Conditional Image Generation (ICCV 2019 · similarity 0.91)
- GAN Memory with No Forgetting (NeurIPS 2020 · similarity 0.89)
- LFS-GAN: Lifelong Few-Shot Image Generation (ICCV 2023 · similarity 0.89)
- Recall-Oriented Continual Learning with Generative Adversarial Meta-Model (AAAI 2024 · similarity 0.86)
- ZeroFlow: Overcoming Catastrophic Forgetting is Easier than You Think (ICML 2025 · 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.
