Research map / Image generation / Autoregressive and latent generation
Towards Conceptual Compression
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 convolutional DRAW instead of a one-pass VAE (Deep VAE) | latent variable image model · CIFAR-10 | NLL bits/dim: 4.54 -> 3.58 | Helped |
| use convolutional DRAW instead of standard DRAW | latent variable image model · CIFAR-10 | NLL bits/dim: 4.13 -> 3.58 | Helped |
| use convolutional DRAW instead of standard DRAW | Omniglot model · Omniglot | NLL nats: 96.5 -> 91.0 | Helped |
| use convolutional DRAW instead of IWAE / RBM | Omniglot model · Omniglot | NLL nats: 103.38 -> 91.0 | Helped |
| compare with autoregressive Pixel RNN | Conv DRAW · CIFAR-10 | NLL bits/dim: 3.00 -> 3.58 | Hurt |
| compare with Pixel RNN | Conv DRAW · ImageNet 32x32 | NLL bits/dim: 3.86 -> 4.40 | Hurt |
| compare with Pixel RNN | Conv DRAW · ImageNet 64x64 | NLL bits/dim: 3.63 -> 4.10 | Hurt |
| scale down input cost (beta < 1) relative to latents (ablation) | Conv DRAW · ImageNet 32x32 samples | sample quality (qualitative) | Helped |
| keep full likelihood objective beta=1 at 64x64 (ablation) | Conv DRAW · ImageNet 64x64 samples | sample coherence (qualitative) | Hurt |
| increase number of DRAW iterations n_t up to 16 (ablation) | Conv DRAW · training speed | wall-clock training time | No clear effect |
| increase iterations beyond n_t=16 (ablation) | Conv DRAW · training speed | wall-clock training | Hurt |
| lower prior variance to zero when generating unstored latents (ablation) | Conv DRAW compression · ImageNet | reconstruction (qualitative) | Mixed |
Closest papers by meaning
- Boosting Generative Image Modeling via Joint Image-Feature Synthesis (NeurIPS 2025 · similarity 0.86)
- When Worse is Better: Navigating the Compression Generation Trade-off In Visual Tokenization (NeurIPS 2025 · similarity 0.86)
- UNICOM: Unified Multimodal Modeling via Compressed Continuous Semantic Representations (NeurIPS 2026 · similarity 0.85)
- Generative Latent Coding for Ultra-Low Bitrate Image Compression (CVPR 2024 · similarity 0.85)
- BiGR: Harnessing Binary Latent Codes for Image Generation and Improved Visual Representation Capabilities (ICLR 2025 · similarity 0.85)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.85)
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.