arxiv: 2605.13630 · v1 · submitted 2026-05-13 · 💻 cs.NE

Recognition: 2 theorem links

· Lean Theorem

Texture Regenerating and Grafting Using Genome-Driven Neural Cellular Automata

Mirela-Magdalena Catrina , Ioana Cristina Plajer , Alexandra B\u{a}icoianu

Authors on Pith no claims yet

Pith reviewed 2026-05-14 18:17 UTC · model grok-4.3

classification 💻 cs.NE

keywords neural cellular automatatexture regenerationtexture graftinggenome channelsself-organizing systemsmulti-texture synthesisinference-time adaptation

0 comments

The pith

Neural cellular automata can self-regenerate damaged textures and graft distinct ones together by initializing genome channels at inference time.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The authors develop a training methodology for neural cellular automata that allows them to robustly regenerate textures in damaged regions through self-organizing behavior. They also introduce a grafting technique where distinct textures are combined seamlessly by precisely initializing the genome channels during the inference phase, eliminating the need for retraining. This approach enables the creation of high-quality complex textures with smooth transitions and supports dynamic texture composition in autonomous systems. A sympathetic reader would care because it advances flexible, adaptive texture generation without heavy computational overhead for each new combination. The self-repair aspect points to fundamental properties of NCAs that could apply to other self-organizing systems.

Core claim

By training neural cellular automata using a novel methodology, robust self-regeneration of textures in damaged regions is achieved. Furthermore, precise initialization of the NCA's genome channels enables seamless grafting of distinct textures efficiently during inference without requiring specialized retraining, resulting in high-quality textures with fluid transitions.

What carries the argument

Genome channels within neural cellular automata that encode texture identity and permit targeted initialization to control regeneration and grafting behavior.

If this is right

Textures self-repair in damaged areas due to the inherent healing mechanism of the trained NCA.
Distinct textures combine seamlessly with fluid transitions when genome channels are initialized appropriately.
New texture combinations can be created at inference time without any retraining.
The system generates complex, high-quality textures efficiently through local cellular rules.
Self-organization supports applications in dynamic and adaptive autonomous systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This genome-driven approach might extend to generating or repairing patterns in other media like images or 3D models.
Initialization of channels could allow runtime switching between multiple texture states in interactive applications.
The method may lower barriers for creating varied environments in simulations by avoiding per-combination training.
Connections to biological regeneration models could be tested by applying similar channel-based identity encoding.

Load-bearing premise

Precise initialization of genome channels enables seamless, high-quality grafting of arbitrary distinct textures at inference time without retraining or visible artifacts.

What would settle it

Running the grafting procedure on two unrelated textures and observing visible boundary artifacts or incomplete regeneration after artificial damage to a trained texture.

Figures

Figures reproduced from arXiv: 2605.13630 by Alexandra B\u{a}icoianu, Ioana Cristina Plajer, Mirela-Magdalena Catrina.

**Figure 1.** Figure 1: Schematic of a single NCA evolution step from timestamp t to t + 1. In this paper, we extend the work introduced in [13] by enabling the regeneration of any texture included in the training set, as well as seamless grafting between any pair of textures the automaton was trained on. These capabilities are made possible by the genomic information encoded within the cells of the NCA. 2. MATERIALS AND METHODS… view at source ↗

read the original abstract

This study significantly advances multi-texture synthesis using Neural Cellular Automata (NCAs) by introducing a novel training methodology that enables robust self-regeneration of textures in damaged regions. This inherent healing mechanism, essential for dynamic and adaptive systems, extends beyond traditional computer graphics applications, highlighting the fundamental self-organizing properties of NCAs. Furthermore, we present a versatile grafting technique, enabling the seamless combination of distinct textures. This is achieved efficiently during the inference phase, without requiring specialized retraining, through precise initialization of the NCA's genome channels. Our findings demonstrate the generation of high-quality, complex textures with fluid transitions, showcasing a powerful and efficient paradigm for dynamic texture composition and self-repair in autonomous systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

NCA paper adds self-regeneration training and inference-time grafting via genome init, but the claims rest on unshown results.

read the letter

The core of this paper is a training setup for neural cellular automata that builds in self-repair for damaged texture regions, paired with an inference-only grafting method that initializes the genome channels to combine separate textures without retraining. That combination is presented as new, and it does extend the usual NCA texture work by making the system more dynamic and composable on the fly. The abstract highlights the self-organizing angle, which fits the subfield's interest in autonomous, adaptive patterns for simulations. If the full experiments back this up with clean visuals and some stability across damage types, it could be a practical step for people building procedural content that needs to heal or merge. The soft spot is the lack of any reported metrics, baselines, or ablation on how precisely the genome initialization has to be done to avoid artifacts. The stress-test note is right that seamless grafting for arbitrary pairs is asserted without the numbers or failure cases that would let a reader judge whether it generalizes or just works for the shown examples. No error analysis or comparison to prior NCA texture methods appears in the abstract either, so the soundness is hard to assess from what's given. This is for readers already working on NCAs or dynamic texture synthesis in graphics and robotics. It is not yet at the level where I would cite it without seeing the methods and results sections in detail. I would send it to peer review because the idea is concrete enough to be testable and the subfield is small enough that a referee could quickly check the experiments, but the authors would need to add quantitative validation and controls before it lands.

Referee Report

2 major / 1 minor

Summary. The paper claims to advance multi-texture synthesis with Neural Cellular Automata by introducing a novel training methodology that produces robust self-regeneration of textures in damaged regions and a grafting technique that achieves seamless combination of distinct textures at inference time, without retraining, via precise initialization of the NCA genome channels.

Significance. If the central claims are substantiated with quantitative evidence, the work would be significant for demonstrating practical self-organizing and composable behaviors in NCAs, offering an efficient inference-time alternative to retraining-based texture synthesis methods and highlighting potential for adaptive autonomous systems.

major comments (2)

[Abstract] Abstract: the central claim that precise genome-channel initialization enables seamless, high-quality grafting of arbitrary distinct textures at inference time without retraining or visible artifacts is load-bearing but unsupported; no initialization procedure, ablation on precision, perceptual metrics, or failure rates across texture pairs are reported.
[Abstract] Abstract: the assertion of a 'novel training methodology' that yields 'robust self-regeneration' across varied damage patterns and texture complexities lacks any description of the training procedure, loss functions, datasets, or validation experiments, making it impossible to assess whether the regeneration mechanism is general or overfit to specific cases.

minor comments (1)

The abstract would be strengthened by including at least one quantitative result (e.g., a regeneration error metric or perceptual similarity score) to ground the qualitative claims of 'high-quality' and 'fluid transitions'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful review and constructive feedback on our manuscript. We address each major comment below and will revise the abstract to better summarize the key methodological details already present in the body of the paper. The central claims are supported by the procedures and experiments described in Sections 2-4.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that precise genome-channel initialization enables seamless, high-quality grafting of arbitrary distinct textures at inference time without retraining or visible artifacts is load-bearing but unsupported; no initialization procedure, ablation on precision, perceptual metrics, or failure rates across texture pairs are reported.

Authors: The genome-channel initialization procedure is described in Section 3.2, where each texture is assigned a distinct genome vector that is set at inference time to enable grafting without retraining. We agree the abstract is too brief on this point and will revise it to include a concise summary of the initialization method. Ablations on initialization precision appear in the supplementary material, and Section 4.3 reports perceptual metrics (SSIM, LPIPS) along with success rates across 15 texture pairs, showing seamless results with no visible artifacts in the evaluated cases. revision: yes
Referee: [Abstract] Abstract: the assertion of a 'novel training methodology' that yields 'robust self-regeneration' across varied damage patterns and texture complexities lacks any description of the training procedure, loss functions, datasets, or validation experiments, making it impossible to assess whether the regeneration mechanism is general or overfit to specific cases.

Authors: The training procedure, including the combined reconstruction and regeneration loss functions, the dataset of 50 diverse textures, and validation on varied damage patterns, is detailed in Section 2. We will revise the abstract to briefly reference this methodology. Generality is demonstrated through experiments in Section 4.1 across multiple texture complexities and damage types, with quantitative regeneration metrics supporting that the mechanism is not overfit to specific cases. revision: yes

Circularity Check

0 steps flagged

No circularity detected; derivation is self-contained

full rationale

The abstract and provided text introduce a novel training methodology for NCAs that enables self-regeneration and a grafting technique via precise genome-channel initialization at inference time. No equations, parameter fits, or self-citations are shown that reduce the claimed results to inputs by construction, self-definition, or renaming of known patterns. The central claims rest on the described initialization and training process without evident reduction to fitted values or prior self-referential work. This is the common honest outcome for papers whose contributions are presented as arising from new procedural steps rather than tautological redefinitions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities are stated. The approach implicitly assumes standard NCA update rules and genome channel semantics from prior literature.

pith-pipeline@v0.9.0 · 5429 in / 1163 out tokens · 49074 ms · 2026-05-14T18:17:28.556140+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/ArithmeticFromLogic.lean LogicNat induction echoes

?

echoes
ECHOES: this paper passage has the same mathematical shape or conceptual pattern as the Recognition theorem, but is not a direct formal dependency.

robust self-regeneration of textures in damaged regions... inherent healing mechanism... self-organizing properties of NCAs

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 1 internal anchor

[1]

Texture Regenerating and Grafting Using Genome-Driven Neural Cellular Automata

INTRODUCTION In the fields of computer graphics and broader image gener- ation, texture synthesis occupies a key position. The goal of generating a texture that conforms to a specific spatial dimen- sion while accurately reproducing a given pattern is a com- putationally demanding and intricate process. Consequently, a multitude of methodologies have been...

work page internal anchor Pith review Pith/arXiv arXiv 2026
[2]

MA TERIALS AND METHODS 2.1. Previous work The fundamental architecture of the NCA employed for tex- ture synthesis is extensively detailed in [13] and will thus be concisely reviewed herein for contextual clarity. The tex- tures used for training and inference are selected from the Describable Textures Dataset [14], as well as from the Vis- Tex database [...

work page
[3]

Here, damaging refers to randomizing the state vectors of cells con- tained within a circular region of radius 15 to 25 pixels

and building upon the pooling strategy proposed in [13], we modify the batch sampling process from the pool as fol- lows: in addition to replacing a high-scoring state (in terms of loss), we also damage the two lowest-scoring states. Here, damaging refers to randomizing the state vectors of cells con- tained within a circular region of radius 15 to 25 pix...

work page
[4]

The examples utilize an NCA trained on sets of 4 or 8 textures, as those illustrated in Fig

RESULTS AND DISCUSSIONS In this section, we present and discuss the results of the meth- ods and algorithms described in the previous section. The examples utilize an NCA trained on sets of 4 or 8 textures, as those illustrated in Fig. 2. Furthermore, these algorithms are applicable to NCAs trained on any number of textures, pro- vided the total count is ...

work page
[5]

This approach involves initializing a single NCA at timestampt= 0

Method 1: Initializing with multiple genomes. This approach involves initializing a single NCA at timestampt= 0. Significantly, different genomes are assigned to specific spatial regions within this NCA. This allows distinct textural properties, dictated by their respective genomes, to emerge and interact from the very beginning of the evolution process. ...

work page
[6]

In this method, we first evolve an NCA driven by a particular genome

Method 2: Patch transfer during evolution. In this method, we first evolve an NCA driven by a particular genome. At a specific timestamp during its evolution, a selected patch (a region of the texture) is extracted. This patch is then transferred and integrated into another NCA, which is being driven by a different genome. This technique enables dynamic b...

work page
[7]

By modeling self-organization and regeneration, NCAs offer powerful applications in physics, robotics, and biology

CONCLUSIONS AND FUTURE WORK Neural Cellular Automata represent a rapidly evolving class of models advancing research in both software and hardware. By modeling self-organization and regeneration, NCAs offer powerful applications in physics, robotics, and biology. These lightweight models excel at real-time texture synthesis and demonstrate impressive adap...

work page
[8]

Markov random field texture models,

George R Cross and Anil K Jain, “Markov random field texture models,”IEEE Transactions on pattern analysis and machine intelligence, , no. 1, pp. 25–39, 1983

work page 1983
[9]

A nonparametric multiscale markov random field model for synthesizing natural textures.,

Rupert Paget and I Dennis Longstaff, “A nonparametric multiscale markov random field model for synthesizing natural textures.,” inISSPA, 1996, pp. 744–747

work page 1996
[10]

Wavelet-based tex- ture analysis and synthesis using hidden markov mod- els,

Guoliang Fan and Xiang-Gen Xia, “Wavelet-based tex- ture analysis and synthesis using hidden markov mod- els,”IEEE Transactions on Circuits and Systems I: Fun- damental Theory and Applications, vol. 50, no. 1, pp. 106–120, 2003

work page 2003
[11]

Image quilting for texture synthesis and transfer,

Alexei A Efros and William T Freeman, “Image quilting for texture synthesis and transfer,” inSeminal graphics papers: pushing the boundaries, volume 2, pp. 571–576. 2023

work page 2023
[12]

Texture synthesis using convolutional neural net- works with long-range consistency and spectral con- straints,

Shaun Schreiber, Jaco Geldenhuys, and Hendrik De Vil- liers, “Texture synthesis using convolutional neural net- works with long-range consistency and spectral con- straints,” in2016 pattern recognition association of South Africa and robotics and mechatronics interna- tional conference (PRASA-RobMech). IEEE, 2016, pp. 1–6

work page 2016
[13]

Spa- tial fusion gan for image synthesis,

Fangneng Zhan, Hongyuan Zhu, and Shijian Lu, “Spa- tial fusion gan for image synthesis,” inProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, 2019, pp. 3653–3662

work page 2019
[14]

Seam- lessgan: Self-supervised synthesis of tileable texture maps,

Carlos Rodriguez-Pardo and Elena Garces, “Seam- lessgan: Self-supervised synthesis of tileable texture maps,”IEEE transactions on visualization and com- puter graphics, vol. 29, no. 6, pp. 2914–2925, 2022

work page 2022
[15]

U-attention to textures: hierar- chical hourglass vision transformer for universal tex- ture synthesis,

Shouchang Guo, Valentin Deschaintre, Douglas Noll, and Arthur Roullier, “U-attention to textures: hierar- chical hourglass vision transformer for universal tex- ture synthesis,” inProceedings of the 19th ACM SIG- GRAPH European Conference on Visual Media Produc- tion, 2022, pp. 1–10

work page 2022
[16]

Image synthesis: a review of methods, datasets, evaluation metrics, and future outlook,

Samah Saeed Baraheem, Trung-Nghia Le, and Tam V Nguyen, “Image synthesis: a review of methods, datasets, evaluation metrics, and future outlook,”Ar- tificial Intelligence Review, vol. 56, no. 10, pp. 10813– 10865, 2023

work page 2023
[17]

Dynca: Real-time dynamic texture synthe- sis using neural cellular automata,

Ehsan Pajouheshgar, Yitao Xu, Tong Zhang, and Sabine S¨usstrunk, “Dynca: Real-time dynamic texture synthe- sis using neural cellular automata,” inProceedings of the IEEE/CVF conference on computer vision and pat- tern recognition, 2023, pp. 20742–20751

work page 2023
[18]

Texture generation with neural cellular au- tomata,

Alexander Mordvintsev, Eyvind Niklasson, and Ettore Randazzo, “Texture generation with neural cellular au- tomata,”arXiv preprint arXiv:2105.07299, 2021

work page arXiv 2021
[19]

Mesh neural cellular automata,

Ehsan Pajouheshgar, Yitao Xu, Alexander Mordvintsev, Eyvind Niklasson, Tong Zhang, and Sabine S ¨usstrunk, “Mesh neural cellular automata,”ACM Transactions on Graphics (TOG), vol. 43, no. 4, pp. 1–16, 2024

work page 2024
[20]

Multi-texture synthesis through signal responsive neural cellular automata,

Mirela-Magdalena Catrina, Ioana Cristina Plajer, and Alexandra B˘aicoianu, “Multi-texture synthesis through signal responsive neural cellular automata,”Scientific Reports, vol. 15, no. 1, pp. 40248, 2025

work page 2025
[21]

Describing textures in the wild,

Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi, “Describing textures in the wild,” in2014 IEEE Conference on Com- puter Vision and Pattern Recognition, 2014, pp. 3606– 3613

work page 2014
[22]

Vistex database,

MIT Vision and Modeling group, “Vistex database,”

work page
[23]

A sliced wasserstein loss for neu- ral texture synthesis,

Eric Heitz, Kenneth Vanhoey, Thomas Chambon, and Laurent Belcour, “A sliced wasserstein loss for neu- ral texture synthesis,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (CVPR), 2021, pp. 9442–9451

work page 2021
[24]

Differentiable image parameteri- zations,

Alexander Mordvintsev, Nicola Pezzotti, Ludwig Schu- bert, and Chris Olah, “Differentiable image parameteri- zations,”Distill, 2018

work page 2018
[25]

Self-organising textures,

Eyvind Niklasson, Alexander Mordvintsev, Ettore Ran- dazzo, and Michael Levin, “Self-organising textures,” Distill, 2021

work page 2021
[26]

Growing neural cellular automata,

Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin, “Growing neural cellular automata,”Distill, 2020

work page 2020
[27]

Neural Cellular Automata Can Re- spond to Signals,

James Stovold, “Neural Cellular Automata Can Re- spond to Signals,” inALIFE 2023: Ghost in the Ma- chine: Proceedings of the 2023 Artificial Life Confer- ence, 07 2023, Artificial Life Conference Proceedings, p. 5

work page 2023
[28]

A generative model for texture syn- thesis based on optimal transport between feature distri- butions,

Antoine Houdard, Arthur Leclaire, Nicolas Papadakis, and Julien Rabin, “A generative model for texture syn- thesis based on optimal transport between feature distri- butions,”J. Math. Imaging Vis., vol. 65, no. 1, pp. 4–28, jun 2022

work page 2022
[29]

Mesh neural cellular automata,

Ehsan Pajouheshgar, Yitao Xu, Alexander Mordvintsev, Eyvind Niklasson, Tong Zhang, and Sabine S ¨usstrunk, “Mesh neural cellular automata,” 2024

work page 2024