Research map / Image generation / Autoregressive and latent generation
Pixel Recurrent Neural Networks
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 a discrete softmax instead of MCGSM for pixel values (ablation) | Row LSTM · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.22 -> 3.06 | Helped |
| Add residual connections without skip connections (ablation) | 12-layer Row LSTM · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.22 -> 3.07 | Helped |
| Add skip connections without residual connections (ablation) | 12-layer Row LSTM · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.22 -> 3.09 | Helped |
| Use both residual and skip connections rather than neither (ablation) | 12-layer Row LSTM · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.22 -> 3.06 | Helped |
| Increase the number of LSTM layers (ablation) | Row LSTM with residual and skip connections · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.30 -> 3.06 | Helped |
| Use Diagonal BiLSTM rather than Row LSTM | PixelRNN · CIFAR-10 test | negative log-likelihood (bits/dim): 3.07 -> 3.00 | Helped |
| Use Row LSTM rather than PixelCNN | Pixel autoregressive model · CIFAR-10 test | negative log-likelihood (bits/dim): 3.14 -> 3.07 | Helped |
| Use Diagonal BiLSTM rather than PixelCNN | Pixel autoregressive model · CIFAR-10 test | negative log-likelihood (bits/dim): 3.14 -> 3.00 | Helped |
| Use Diagonal BiLSTM rather than Row LSTM | PixelRNN · MNIST test | negative log-likelihood (nats): 80.54 -> 79.20 | Helped |
| Increase the Row LSTM from 1 to 12 layers with residual and skip connections (ablation) | Row LSTM · CIFAR-10 validation | negative log-likelihood (bits/dim): 3.30 -> 3.06 | Helped |
Closest papers by meaning
- Learning Implicitly Recurrent CNNs Through Parameter Sharing (ICLR 2019 · similarity 0.89)
- PixelCNN Models with Auxiliary Variables for Natural Image Modeling (ICML 2017 · similarity 0.88)
- PixelCNN++: Improving the PixelCNN with Discretized Logistic Mixture Likelihood and Other Modifications (ICLR 2017 · similarity 0.88)
- Rethinking Generative Image Pretraining: How Far Are We From Scaling Up Next-Pixel Prediction? (ICML 2026 · similarity 0.88)
- DeepStereo: Learning to Predict New Views From the World's Imagery (CVPR 2016 · similarity 0.88)
- Recurrent neural network dynamical systems for biological vision (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.