Show HN: High-Res Neural Cellular Automata
TL;DR Highlight
A coarse NCA plus a tiny decoder renders real-time output at any resolution.
Who Should Read
A web or graphics developer interested in generative graphics and texture synthesis
Core Mechanics
- Resolution was the NCA bottleneck
- Neural Cellular Automata (NCAs) apply the same learned local rule in every cell, over and over, to build patterns.
- Until now they stayed at low resolution.
- Three causes limit resolution
- Training time and memory grow quadratically with grid size.
- Information only moves between neighbors, so distant cells struggle to communicate.
- Real-time inference at high resolution is compute-heavy.
- A coarse-grid NCA plus a small decoder
- The NCA runs only on a coarse grid.
- A lightweight implicit decoder (LPPN) takes cell states and local coordinates and outputs attributes like color and surface normal.
- One model, any output resolution
- Both the decoder and NCA updates are local, so it parallelizes well.
- The authors report real-time high-res output on 2D grids, 3D grids, and meshes.
- NCA and LPPN train together
- A point inside a triangle is described by interpolated vertex cell states plus a local coordinate.
- Both go into the shared MLP (LPPN), trained end-to-end with the NCA.
- Custom losses for high-res supervision
- There are separate losses for growth from a seed and for texture synthesis.
- The authors say extra memory and compute are minimal.
- Self-organizing traits like regeneration and robustness are preserved.
Community Reactions
- A user reported the image collapses under heavy brushing
- A commenter said five horizontal swipes over a stabilized image destroyed the kitten.
- They asked whether that is expected.
- An opinion that it resembles texture sampling
- One commenter thought it works like iterative texture sampling, memorizing the image through local rules instead of coordinate lookup.
- A pointer to the distill article for beginners
- A commenter recommended the earlier distill.pub growing-ca article as a clearer intro.
- A WebGPU implementation was shared
- A commenter built a WebGPU version of the original growing neural CA and posted the repo.
- A usability gripe about the demo
- A commenter said switching Target Image in the Growing Demo resets the canvas, so they couldn't stamp one image over another.
How to Apply
- Play with the public demo
- Open cells2pixels.github.io today and erase parts of an image with the brush in the Growing demo.
- Watch whether it regrows to see the regeneration behavior.
- Change LPPN Scale and compare
- Switch LPPN Scale (e.g., x4) and see how the same cell states render.
- Check that it still runs in real time at higher resolution.
- Sketch a coarse-state plus decoder design
- If an expensive high-res step bottlenecks your generation pipeline, sketch a split into a low-res state and a light decoder.
- The source only covers graphics, so validate with a small prototype first.
Terminology
Related Papers
We're releasing the weights of Mistral Large 4
Mistral released Large 4 weights, claiming top open-weight scores on coding benchmarks
Gemini 4 Carbon is the new #1 on the Artificial Analysis Intelligence Index v5
Gemini 4 Carbon takes #1 on Intelligence Index v5 with 59 and is second in coding agents.
Gemini 4 Carbon
Google unveils Gemini 4 Carbon, a cybersecurity-focused model, but access is limited to select cyber defenders.
Dayjob: Tax – evaluating AI agents on 50 professional tax assignments
On 50 professional tax tasks, the best AI agent passed only 10.8%
Anthropic S-1
Anthropic filed an S-1 to list Class A shares on Nasdaq as 'ANT'.
A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
A fine-tuned Inception v3 beat GPT-6 Astra and Claude Fable 5.1 on 200 skin lesions
Why Does Structured Output Fail? An Empirical Study of Prompt-Enforced Pydantic Schemas Across Nine LLMs
Breaking structured-output failures into types showed markdown fences as the top cause, often hiding length violations.
America has lost 200k of jobs in media, film and the arts. How much is AI?
US creative-industry jobs fell by over 200k in four years, but AI's share of the blame is hard to prove