Cells2Pixels¶

Development line: project:cells2pixels · thread cells2pixels-development
Last event: - · 0 dated since - · Researched: 2026-09-05 · confidence: high

What it is¶

Cells2Pixels is the code and interactive demo of Neural Cellular Automata: From Cells to Pixels for generative graphics researchers.

• NCA and Local Pattern Producing Network (LPPN) train jointly.
• Core modes cover 2D morphogenesis, PBR textures, textures on meshes, and 3D textures.
• Experimental modes cover dynamic textures, radiance field, and voxel growth.

The repository is built for training, without a packaged release or ready Colab. The README still lists Colab and testing as TODO items. It fits reproducible NCA experiments, not production image generation.

Development line¶

• The dated line is not written up yet; what is known stands in the sections below.

What changed¶

2026-07-11 — Cells2Pixels arrived as a SIGGRAPH 2026 implementation. A coarse cellular grid evolves through local updates. The LPPN decodes interpolated cell states and local coordinates into images or surface properties at arbitrary resolution.

How to use this¶

As of 2026-07-11, treat Cells2Pixels as a project requiring source review through its website and repository; no practitioner workflow change is justified from the dated links alone.

1. Install dependencies with pip install -r requirements.txt. Kaolin is optional for runs without meshes. — https://github.com/TheDevilWillBeBee/Cells2Pixels
2. Download datasets and input directory structures with python scripts/download_data.py. — https://github.com/TheDevilWillBeBee/Cells2Pixels
3. Run training with python train.py --config <path>. Start with configs/nca2d/growing.yaml for 2D growth, configs/nca2d/pbr_texture.yaml for PBR, configs/meshnca/texture.yaml for meshes, or configs/nca3d/3d_texture.yaml for 3D. — https://github.com/TheDevilWillBeBee/Cells2Pixels
4. Pass --test after training to load the checkpoint and save images and rollout videos into the configured outputs directory. — https://github.com/TheDevilWillBeBee/Cells2Pixels
5. Test model behavior in the interactive canvas by choosing a mode, adjusting the LPPN scale, and clicking or touching the NCA. — https://cells2pixels.github.io/

Best practices¶

• Match PyTorch and Kaolin versions with requirements. Kaolin is needed only for mesh rendering and rasterization. — https://github.com/TheDevilWillBeBee/Cells2Pixels
• Do not present dynamic texture, radiance-field growth, or voxel growth as paper results. The README separates these experimental modes from the four paper training modes. — https://github.com/TheDevilWillBeBee/Cells2Pixels
• Evaluate models with a saved checkpoint and an explicit test rollout instead of relying on the training run alone. — https://github.com/TheDevilWillBeBee/Cells2Pixels

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• The project page and README provide no dated repository changelog, so publication dates for code, demos, and specific configs cannot be verified.
• No confirmed packaged release, checkpoint catalogue, or supported production workflow exists.

Sources¶

Agent brief¶

• Subject: project:cells2pixels, thread cells2pixels-development, 0 dated events - → -.
• Practical note: As of 2026-07-11, treat Cells2Pixels as a project requiring source review through its website and repository; no practitioner workflow change is justified from the dated links alone.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.