Research map / Probabilistic and scientific modeling / Variational and generative sampling
Variational Neural Cellular Automata
Rasmus Berg Palm, Miguel González Duque, Shyam Sudhakaran, Sebastian Risi
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 |
|---|---|---|---|
| Use the doubling VNCA decoder instead of the deep convolutional PixelCNN decoder | Binarized MNIST · Binarized MNIST test | log likelihood (nats): -84.23 vs -81.3 | Hurt |
| Set β=100 to prioritize better samples over reconstructions (ablation) | VNCA · CelebA 64×64 test | bits per dimension: 4.42 -> 5.40 | Hurt |
| Remove doubling layers and train with pool-based damage recovery (ablation) | VNCA · MNIST | log likelihood (nats): -84.23 -> -90.97 | Hurt |
| Remove doubling layers and train with pool-based damage recovery (ablation) | VNCA · CelebA 32×32 | bits per dimension: 5.03 bits per dimension | Hurt |
| Increase update-network depth or width (ablation) | VNCA · VNCA model variations | generative modeling performance | No clear effect |
| Increase latent-space size to 512 (ablation) | VNCA · VNCA model variations | generative modeling performance | No clear effect |
| Increase NCA hidden-state size to 512 (ablation) | VNCA · VNCA model variations | generative modeling performance | No clear effect |
| Use L=1 rather than L=4 logistic mixture components (ablation) | VNCA · VNCA model variations | generative modeling performance | Improved |
| Replace 1×1 convolution layers with 3×3 convolutions (ablation) | VNCA · VNCA model variations | generative modeling performance | No clear effect |
| Use a separate update network at each scale (ablation) | VNCA · VNCA model variations | generative modeling performance | Improved |
| Reduce batch size to 4 (ablation) | VNCA · VNCA training | training quality | Hurt |
| Replace the update function with a GRU (ablation) | VNCA · VNCA model variations | generative modeling performance | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- PixelCNN — PixelCNN++: Improving the PixelCNN with Discretized Logistic Mixture Likelihood and Other Modifications (ICLR 2017) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- [Re] Variational Neural Cellular Automata (NeurIPS 2023 · similarity 0.88)
- eXponential FAmily Dynamical Systems (XFADS): Large-scale nonlinear Gaussian state-space modeling (NeurIPS 2024 · similarity 0.83)
- Deep Neural Cellular Potts Models (ICML 2025 · similarity 0.83)
- CAX: Cellular Automata Accelerated in JAX (ICLR 2025 · similarity 0.83)
- Brain-like Variational Inference (NeurIPS 2025 · similarity 0.83)
- A solvable model of learning generative diffusion: theory and insights (NeurIPS 2025 · similarity 0.82)
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.
