Research map / Image generation / Fast diffusion sampling
Anytime Sampling for Autoregressive Models via Ordered Autoencoding
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 ordered rather than unordered latent codes | VQ-VAE anytime sampling · CIFAR-10 and CelebA | FID | Helped |
| Use ordered rather than unordered latent codes | VQ-VAE anytime sampling · CIFAR-10 and CelebA | sample quality as code length increases | Helped |
| Use channel-wise rather than spatial-wise quantization (ablation) | OVQ-VAE · MNIST and CelebA | reconstruction quality with truncated codes | Helped |
| Use a Transformer or LSTM rather than PixelCNN (ablation) | Autoregressive model on ordered channel-wise codes · CIFAR-10 | FID | Helped |
| Use PixelCNN rather than a 1D-sequence autoregressive model (ablation) | Autoregressive model on ordered channel-wise codes · CIFAR-10, 0.2 fraction of full code length | FID | Mixed |
| Use Geo(0.03) rather than the uniform masking distribution (ablation) | OVQ-VAE order-inducing objective · CIFAR-10 | FID | Hurt |
| Use Geo(0.1) rather than the uniform masking distribution (ablation) | OVQ-VAE order-inducing objective · CIFAR-10, 0.6/0.8/1.0 code fractions | FID | Hurt |
| Train a standard VAE with the order-inducing objective (ablation) | Standard VAE · CIFAR-10 | reconstruction quality with truncated codes | Helped |
| Train the autoregressive model on PCA-projected continuous data rather than raw data (ablation) | Transformer · MNIST | bits/dim: 0.80 -> 1.22 | Hurt |
| Use ordered rather than unordered codes (ablation) | VQ-VAE reconstruction · CIFAR-10 | reconstruction error under aggressive truncation | Helped |
| Use ordered rather than unordered codes (ablation) | VQ-VAE autoregressive sampling · CIFAR-10, full code length | FID | Helped |
| Use anytime sampling rather than sampling the full code | OVQ-VAE · CelebA | sampling runtime: 40% | Helped |
Closest papers by meaning
- Stabilize the Latent Space for Image Autoregressive Modeling: A Unified Perspective (NeurIPS 2024 · similarity 0.87)
- Distilled Decoding 1: One-step Sampling of Image Auto-regressive Models with Flow Matching (ICLR 2025 · similarity 0.87)
- Towards Better & Faster Autoregressive Image Generation: From the Perspective of Entropy (NeurIPS 2025 · similarity 0.87)
- Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI (NeurIPS 2024 · similarity 0.87)
- NFIG: Multi-Scale Autoregressive Image Generation via Frequency Ordering (NeurIPS 2025 · similarity 0.86)
- Autoregressive Image Generation without Vector Quantization (NeurIPS 2024 · 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.