Research map / Probabilistic and scientific modeling / Variational and generative sampling
The Hessian Penalty: A Weak Prior for Unsupervised Disentanglement
William Peebles, John Peebles, Jun-Yan Zhu, Alexei Efros, Antonio Torralba
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 |
|---|---|---|---|
| fine-tune ProGAN with the Hessian Penalty | ProGAN · Edges+Shoes | PPL: 1809.7 -> 554.1 | Improved |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · CLEVR-Simple | PPL: 61.5 -> 39.7 | Improved |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · CLEVR-Complex | PPL: 92.8 -> 74.7 | Improved |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · CLEVR-U | PPL: 720.2 -> 61.6 | Improved |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · CLEVR-1FOV | PPL: 35.5 -> 10.0 | Improved |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · Edges+Shoes | FID: 14.0 -> 17.3 | Hurt |
| fine-tune ProGAN with the Hessian Penalty | ProGAN · CLEVR-Simple | FID: 3.5 -> 6.1 | Hurt |
| train ProGAN from scratch with the Hessian Penalty instead of fine-tuning | ProGAN + Hessian Penalty · CLEVR-Simple | FID: 6.1 -> 25.0 | Hurt |
| train with the Hessian Penalty | ProGAN · CLEVR-1FOV | FID: 11.5 -> 2.3 | Improved |
| apply Hessian Penalty to ProGAN | ProGAN · CLEVR-Simple | D% (diagonal max): 69.8% -> 85.7% | Improved |
| apply Hessian Penalty to ProGAN with underparameterized latent | ProGAN · CLEVR-U | D_R: 2.3 -> 2.2 | No clear effect |
| apply Hessian Penalty only to early layers (ablation) | ProGAN · generator layers | disentanglement | Hurt |
Closest papers by meaning
- Orthogonal Jacobian Regularization for Unsupervised Disentanglement in Image Generation (ICCV 2021 · similarity 0.93)
- Disentanglement via Latent Quantization (NeurIPS 2023 · similarity 0.90)
- Hyperprior Induced Unsupervised Disentanglement of Latent Representations (AAAI 2019 · similarity 0.90)
- Disentanglement as Identifiable Pushforward Factorisation (NeurIPS 2026 · similarity 0.88)
- Tripod: Three Complementary Inductive Biases for Disentangled Representation Learning (ICML 2024 · similarity 0.87)
- Investigating the Overlooked Hessian Structure: From CNNs to LLMs (ICML 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.
