Research map / Image generation / Autoregressive and latent generation
Learning Disentangled Representation by Exploiting Pretrained Generative Models: A Contrastive Learning View
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 DisCo with a pretrained StyleGAN2 instead of CF | Disentangled representation learning · Shapes3D | MIG: 0.307 \pm 0.124 -> 0.512 \pm 0.068 | Helped |
| Use DisCo with a pretrained StyleGAN2 instead of CF | Disentangled representation learning · Shapes3D | DCI: 0.525 \pm 0.078 -> 0.708 \pm 0.048 | Helped |
| Add entropy-based domination loss (ablation) | DisCo with hard-negative flipping · Shapes3D | MIG: 0.134 -> 0.547 | Helped |
| Add entropy-based domination loss (ablation) | DisCo with hard-negative flipping · Shapes3D | DCI: 0.681 -> 0.730 | Helped |
| Flip hard negatives into positives (ablation) | DisCo with entropy-based domination loss · Shapes3D | DCI: 0.627 -> 0.730 | Helped |
| Represent image changes as differences in encoded representations instead of concatenation (ablation) | DisCo · Shapes3D | MIG: 0.023 -> 0.562 | Helped |
| Use contrastive learning instead of classification (ablation) | DisCo variation space · Shapes3D | MIG: 0.002 -> 0.562 | Helped |
| Use the W latent space rather than Z (ablation) | DisCo with unit-length navigator matrix · Shapes3D | MIG: 0.242 -> 0.547 | Helped |
| Use an orthonormal navigator matrix instead of a unit-length-column matrix in W (ablation) | DisCo · Shapes3D | MIG: 0.547 -> 0.551 | No clear effect |
| Use 3 fully-connected layers for the navigator instead of a unit-length-column matrix in W (ablation) | DisCo · Shapes3D | MIG: 0.547 -> 0.340 | Hurt |
| Use a ResNet-50 encoder instead of the lightweight encoder (ablation) | DisCo · Shapes3D | MIG: 0.562 -> 0.03 | Hurt |
| Set hard-negative threshold below 0.9 (ablation) | DisCo · Shapes3D | MIG: 0.508 -> 0.244 | Hurt |
Closest papers by meaning
- Transferring disentangled representations: bridging the gap between synthetic and real images (NeurIPS 2024 · similarity 0.91)
- Disentangled Representation Learning via Flow Matching (NeurIPS 2026 · similarity 0.90)
- Explicit Disentanglement of Appearance and Perspective in Generative Models (NeurIPS 2019 · similarity 0.89)
- On Causally Disentangled Representations (AAAI 2022 · similarity 0.89)
- DisenBooth: Identity-Preserving Disentangled Tuning for Subject-Driven Text-to-Image Generation (ICLR 2024 · similarity 0.88)
- Synergy Between Sufficient Changes and Sparse Mixing Procedure for Disentangled Representation Learning (ICLR 2025 · similarity 0.88)
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.