pith. machine review for the scientific record. sign in

arxiv: 2604.13730 · v1 · submitted 2026-04-15 · 💻 cs.CV

Recognition: unknown

ReConText3D: Replay-based Continual Text-to-3D Generation

Muhammad Ahmed Ullah Khan , Muhammad Haris Bin Amir , Didier Stricker , Muhammad Zeshan Afzal
Authors on Pith no claims yet

Pith reviewed 2026-05-10 13:47 UTC · model grok-4.3

classification 💻 cs.CV
keywords continual learningtext-to-3D generationreplay memoryk-center clusteringcatastrophic forgettingincremental learning3D generationToys4K-CL
0
0 comments X

The pith

ReConText3D uses a compact replay memory of text embeddings to let text-to-3D models learn new categories while retaining old generation ability.

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

Standard text-to-3D models lose the ability to generate previously seen 3D assets when trained on new textual categories one after another. The paper shows this forgetting can be reduced by keeping a small replay buffer whose members are chosen through k-center clustering on text embeddings, then rehearsing them during new training. The approach requires no changes to the underlying generative model and is tested on a new class-incremental benchmark called Toys4K-CL. A reader would care because it suggests 3D assets can be added to a generative system over time as new text descriptions become available, rather than requiring full retraining from scratch each time new objects appear.

Core claim

ReConText3D is the first framework for continual text-to-3D generation. It first shows that existing text-to-3D models suffer catastrophic forgetting when trained incrementally on new categories. The method then constructs a compact and diverse replay memory by applying k-center selection directly to text embeddings, enabling the model to learn new 3D categories from textual descriptions while preserving synthesis quality for prior assets. Experiments on the Toys4K-CL benchmark, built from balanced and semantically diverse splits of the Toys4K dataset, demonstrate consistent outperformance over baselines across multiple generative backbones.

What carries the argument

Compact replay memory formed by k-center clustering on text embeddings, which selects representative prior examples for rehearsal during new class training.

If this is right

  • Generative models can be updated with new textual descriptions of 3D objects over time without complete retraining.
  • Generation quality for both previously seen and newly added classes stays high on balanced incremental splits.
  • The same replay selection works across different text-to-3D backbones without any model-specific modifications.
  • A dedicated benchmark now exists for measuring how well continual learning methods perform in the text-to-3D setting.

Where Pith is reading between the lines

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

  • The success of text-only selection for replay suggests that semantic information in embeddings is already sufficient to capture the diversity needed for 3D shape rehearsal.
  • Similar replay buffers could be tested in related generation tasks such as text-to-image or text-to-video where forgetting is also observed.
  • Future extensions might combine the buffer with lightweight regularization or dynamic buffer sizing to further reduce storage while preserving performance.
  • This approach could support practical systems that grow libraries of 3D assets by adding categories as new descriptive text becomes available.

Load-bearing premise

That a compact replay buffer selected solely by k-center clustering on text embeddings is sufficient to rehearse and retain the full generative capability for prior classes without any architectural changes or additional regularization.

What would settle it

If incremental training on new classes using only the k-center text-embedding replay buffer still produces large drops in generation quality for old classes on the Toys4K-CL splits, or if models without the replay show no forgetting at all, the central claim would not hold.

Figures

Figures reproduced from arXiv: 2604.13730 by Didier Stricker, Muhammad Ahmed Ullah Khan, Muhammad Haris Bin Amir, Muhammad Zeshan Afzal.

Figure 1
Figure 1. Figure 1: Top: The Toys4K-CL benchmark splits data into base classes and novel classes with adversarial splits and semantic diversity while maintaining a comparable number of classes and total assets across both stages. Bottom: Continual text-to-3D generation. Naive fine-tuning on novel data, in this case dog, leads to catastrophic forgetting on base class horse, while continual learning mitigates forgetting, preser… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the ReConText3D framework. Base captions are encoded with CLIP and selected via k-center sampling under a count-aware budget to form replay memory R, which is combined with novel data to incrementally train GθB∪N while preserving base￾class synthesis. Continual Learning. Continual learning aims to develop models capable of learning sequentially while mitigating catastrophic forgetting. Major st… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparision of continual text-to-3D baselines on base class assets: From left to right: Ground Truth, Base Training, Fine-tuning, L2-SP, and our method. Results on TRELLIS are shown on the top, while Shap-E is shown on the bottom. v3 [39] and PointNet++ [28] feature embeddings, for ap￾pearance and geometric quality, respectively. Forgetting (F) is measured as the relative percentage drop in bas… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on replay memory size. Performance of our replay strategy under varying replay memory budgets/sizes on TRELLIS-XL and Shap-E backbones. ample (row 1), it generates a lizard, which is a novel class, instead of the original dragon class from the base set. Re￾ConText3D, on the contrary, retains semantic understanding and structural detail for old classes. Additionally, as seen in [PITH_FULL_IMAGE:fi… view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparision of novel class assets across baselines. Broader object domains may also introduce challenges such as replay-memory saturation and semantic drift over longer training horizons. Moreover, current text-to-3D evaluations still rely on 2D-adapted metrics, future work should there￾fore explore perceptual and physical realism metrics defined directly in 3D space for more faithful assessmen… view at source ↗
Figure 7
Figure 7. Figure 7: Class distribution statistics for the Toys4K-CL benchmark (train and test). Both splits exhibit similar long-tailed behaviour. close to the fine-tuning baseline, indicating that replay does not overly constrain plasticity. This pattern is consistent across both backbones and across both simple and visually complex categories (such as chess piece, glass, radio), sup￾porting our claim that semantic replay ca… view at source ↗
Figure 8
Figure 8. Figure 8: Number of base class assets (replay examplers) returned [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Qualitative comparison of continual text-to-3D baselines against ReConText3D (Ours) on [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative comparison of continual text-to-3D baselines against ReConText3D (Ours) on [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualitative comparison of continual text-to-3D baselines against ReConText3D (Ours) on [PITH_FULL_IMAGE:figures/full_fig_p018_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative comparison of continual text-to-3D baselines against ReConText3D (Ours) on [PITH_FULL_IMAGE:figures/full_fig_p019_12.png] view at source ↗
read the original abstract

Continual learning enables models to acquire new knowledge over time while retaining previously learned capabilities. However, its application to text-to-3D generation remains unexplored. We present ReConText3D, the first framework for continual text-to-3D generation. We first demonstrate that existing text-to-3D models suffer from catastrophic forgetting under incremental training. ReConText3D enables generative models to incrementally learn new 3D categories from textual descriptions while preserving the ability to synthesize previously seen assets. Our method constructs a compact and diverse replay memory through text-embedding k-Center selection, allowing representative rehearsal of prior knowledge without modifying the underlying architecture. To systematically evaluate continual text-to-3D learning, we introduce Toys4K-CL, a benchmark derived from the Toys4K dataset that provides balanced and semantically diverse class-incremental splits. Extensive experiments on the Toys4K-CL benchmark show that ReConText3D consistently outperforms all baselines across different generative backbones, maintaining high-quality generation for both old and new classes. To the best of our knowledge, this work establishes the first continual learning framework and benchmark for text-to-3D generation, opening a new direction for incremental 3D generative modeling. Project page is available at: https://mauk95.github.io/ReConText3D/.

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.

Referee Report

2 major / 1 minor

Summary. The paper presents ReConText3D, the first framework for continual text-to-3D generation. It shows that existing text-to-3D models suffer catastrophic forgetting under incremental training, then proposes a replay-based approach that builds a compact replay buffer via k-center clustering on text embeddings to rehearse prior classes without any architectural modifications or extra regularization. A new benchmark Toys4K-CL is introduced with balanced class-incremental splits from Toys4K, and experiments across generative backbones demonstrate consistent outperformance on both old and new classes.

Significance. If the central empirical claims hold, the work is significant as the first systematic treatment of continual learning for text-to-3D generation. It contributes a new benchmark (Toys4K-CL) and demonstrates that a simple, architecture-agnostic replay strategy can mitigate forgetting in this domain. Credit is due for the reproducible benchmark construction and the empirical validation across multiple backbones; these elements provide a concrete foundation for future incremental 3D generative modeling research.

major comments (2)
  1. [Method] Method (replay memory construction via text-embedding k-Center selection): the central claim that a small buffer of texts chosen solely by k-center clustering in CLIP-style embedding space is sufficient to retain full prior-class generative capability is load-bearing. Because the underlying text-to-3D model maps text to 3D geometry/appearance via score distillation or latent optimization, semantic text embeddings can group prompts that produce visually dissimilar 3D assets; rehearsal on only the selected texts therefore risks under-sampling the learned output manifold. The manuscript should include an ablation that isolates the k-center component (e.g., random text selection or storing a few generated latents) to verify that the reported retention of old-class quality is not an artifact of the specific Toys4K-CL splits.
  2. [Experiments] Experiments section (Toys4K-CL results): while consistent outperformance across backbones is reported, the manuscript provides no details on run-to-run variance, statistical significance tests, or confidence intervals. In addition, no ablation specifically removes or replaces the k-center selection step itself. These omissions make it difficult to assess whether the gains are robust or merely plausible on the chosen benchmark splits.
minor comments (1)
  1. [Abstract] Abstract: the description of Toys4K-CL would benefit from a brief statement of the number of classes, the incremental split sizes, and the precise metrics used for old-class retention.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below. We agree that additional ablations and statistical reporting will strengthen the paper and will incorporate them in the revised manuscript.

read point-by-point responses

  1. Referee: [Method] Method (replay memory construction via text-embedding k-Center selection): the central claim that a small buffer of texts chosen solely by k-center clustering in CLIP-style embedding space is sufficient to retain full prior-class generative capability is load-bearing. Because the underlying text-to-3D model maps text to 3D geometry/appearance via score distillation or latent optimization, semantic text embeddings can group prompts that produce visually dissimilar 3D assets; rehearsal on only the selected texts therefore risks under-sampling the learned output manifold. The manuscript should include an ablation that isolates the k-center component (e.g., random text selection or storing a few generated latents) to verify that the reported retention of old-class quality is not an artifact of the specific Toys4K-CL splits.

    Authors: We acknowledge the referee's concern that CLIP-style text embeddings may not perfectly align with visual dissimilarities in generated 3D assets. Our design choice is grounded in the observation that text-to-3D models are conditioned directly on these embeddings, so selecting diverse prompts in embedding space provides a principled way to cover the input distribution without architecture-specific modifications. To directly address the risk of under-sampling, we will add a new ablation in the revised manuscript that compares k-center selection against random text selection and, where feasible given memory constraints, against replay of generated latents. This will confirm that the reported retention on Toys4K-CL is attributable to the k-center strategy rather than benchmark-specific artifacts. revision: yes

  2. Referee: [Experiments] Experiments section (Toys4K-CL results): while consistent outperformance across backbones is reported, the manuscript provides no details on run-to-run variance, statistical significance tests, or confidence intervals. In addition, no ablation specifically removes or replaces the k-center selection step itself. These omissions make it difficult to assess whether the gains are robust or merely plausible on the chosen benchmark splits.

    Authors: We agree that the absence of variance reporting and an explicit ablation of the k-center component limits the ability to assess robustness. In the revised manuscript we will report all main results as means over multiple independent runs with standard deviations and 95% confidence intervals. We will also add statistical significance tests (e.g., paired t-tests against baselines) for key metrics. In addition, we will include a dedicated ablation that replaces k-center selection with random text sampling to isolate its contribution on the Toys4K-CL splits. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical method validated on new benchmark

full rationale

The paper proposes ReConText3D as an empirical framework for continual text-to-3D generation: it applies standard k-center clustering on text embeddings to build a replay buffer, demonstrates catastrophic forgetting in existing models, and reports outperformance on the introduced Toys4K-CL benchmark across generative backbones. No load-bearing step reduces by construction to a fitted parameter, self-definition, or self-citation chain; the replay heuristic is presented as a practical choice and evaluated externally via experiments rather than derived tautologically from its own inputs. The derivation chain consists of method description plus benchmark results, remaining self-contained against external validation.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 0 invented entities

The approach rests on the assumption that text embeddings capture sufficient semantic diversity for replay selection and that standard generative backbones can be trained incrementally with rehearsal alone; no new physical entities or mathematical axioms are introduced.

free parameters (1)
  • k (number of centers in k-Center selection)
    The size of the replay buffer is a hyper-parameter that controls memory size versus retention quality; its value is chosen per experiment but not derived.

pith-pipeline@v0.9.0 · 5548 in / 1246 out tokens · 28738 ms · 2026-05-10T13:47:49.979166+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

54 extracted references · 14 canonical work pages · 6 internal anchors

  1. [1]

    Foundation model-powered 3d few-shot class in- cremental learning via training-free adaptor

    Sahar Ahmadi, Ali Cheraghian, Morteza Saberi, Md Towsif Abir, Hamidreza Dastmalchi, Farookh Hussain, and Shafin Rahman. Foundation model-powered 3d few-shot class in- cremental learning via training-free adaptor. InProceedings of the Asian Conference on Computer Vision, pages 2282– 2299, 2024. 3

    2024

  2. [2]

    Building Normalizing Flows with Stochastic Interpolants

    Michael S Albergo and Eric Vanden-Eijnden. Building nor- malizing flows with stochastic interpolants.arXiv preprint arXiv:2209.15571, 2022. 2

    work page internal anchor Pith review arXiv 2022

  3. [3]

    End-to-end incremental learning

    Francisco Manuel Castro Pay ´an, Manuel J Mar ´ın-Jim´enez, Nicol´as Guil-Mata, Cordelia Schmid, Karteek Alahari, et al. End-to-end incremental learning. 2018. 3

    2018

  4. [4]

    Riemannian walk for incremen- tal learning: Understanding forgetting and intransigence

    Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajan- than, and Philip HS Torr. Riemannian walk for incremen- tal learning: Understanding forgetting and intransigence. In Proceedings of the European conference on computer vision (ECCV), pages 532–547, 2018. 2, 3

    2018

  5. [5]

    Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation

    Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fan- tasia3d: Disentangling geometry and appearance for high- quality text-to-3d content creation. InProceedings of the IEEE/CVF international conference on computer vision, pages 22246–22256, 2023. 2

    2023

  6. [6]

    3dtopia-xl: Scaling high- quality 3d asset generation via primitive diffusion

    Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high- quality 3d asset generation via primitive diffusion. InPro- ceedings of the Computer Vision and Pattern Recognition Conference, pages 26576–26586, 2025. 2

    2025

  7. [7]

    Learning without mem- orizing

    Prithviraj Dhar, Rajat Vikram Singh, Kuan-Chuan Peng, Ziyan Wu, and Rama Chellappa. Learning without mem- orizing. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5138–5146,

  8. [8]

    I3dol: Incremental 3d object learning without catas- trophic forgetting

    Jiahua Dong, Yang Cong, Gan Sun, Bingtao Ma, and Lichen Wang. I3dol: Incremental 3d object learning without catas- trophic forgetting. InProceedings of the AAAI conference on artificial intelligence, pages 6066–6074, 2021. 3

    2021

  9. [9]

    How to continually adapt text-to-image diffusion models for flexible customization?Advances in Neural In- formation Processing Systems, 37:130057–130083, 2024

    Jiahua Dong, Wenqi Liang, Hongliu Li, Duzhen Zhang, Meng Cao, Henghui Ding, Salman H Khan, and Fahad Shah- baz Khan. How to continually adapt text-to-image diffusion models for flexible customization?Advances in Neural In- formation Processing Systems, 37:130057–130083, 2024. 3

    2024

  10. [10]

    Learning a unified classifier incrementally via rebalancing

    Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 831–839,

  11. [11]

    GPT-4o System Card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perel- man, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli- hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    A survey on text-to-3d contents generation in the wild.arXiv preprint arXiv:2405.09431, 2024

    Chenhan Jiang. A survey on text-to-3d contents generation in the wild.arXiv preprint arXiv:2405.09431, 2024. 2

    work page arXiv 2024

  13. [13]

    Shap-e: Generating conditional 3d implicit functions

    Heewoo Jun and Alex Nichol. Shap-e: Generat- ing conditional 3d implicit functions.arXiv preprint arXiv:2305.02463, 2023. 2, 4, 6, 1

    work page arXiv 2023

  14. [14]

    Generative ai meets 3d: A survey on text-to-3d in aigc era.arXiv preprint arXiv:2305.06131, 2023

    Chenghao Li, Chaoning Zhang, Joseph Cho, Atish Wagh- wase, Lik-Hang Lee, Francois Rameau, Yang Yang, Sung- Ho Bae, and Choong Seon Hong. Generative ai meets 3d: A survey on text-to-3d in aigc era.arXiv preprint arXiv:2305.06131, 2023. 2

    work page arXiv 2023

  15. [15]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelli- gence, 40(12):2935–2947, 2017. 2, 3

    2017

  16. [16]

    Luciddreamer: Towards high- fidelity text-to-3d generation via interval score matching

    Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiao- gang Xu, and Yingcong Chen. Luciddreamer: Towards high- fidelity text-to-3d generation via interval score matching. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 6517–6526, 2024. 2

    2024

  17. [17]

    Magic3d: High-resolution text-to-3d content creation

    Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3d: High-resolution text-to-3d content creation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 300–309, 2023. 2

    2023

  18. [18]

    Flow Matching for Generative Modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximil- ian Nickel, and Matt Le. Flow matching for generative mod- eling.arXiv preprint arXiv:2210.02747, 2022. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  19. [19]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022. 2

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  20. [20]

    Scalable 3d captioning with pretrained models

    Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36: 75307–75337, 2023. 6

    2023

  21. [21]

    Packnet: Adding mul- tiple tasks to a single network by iterative pruning

    Arun Mallya and Svetlana Lazebnik. Packnet: Adding mul- tiple tasks to a single network by iterative pruning. InPro- ceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 7765–7773, 2018. 3

    2018

  22. [22]

    Piggy- back: Adapting a single network to multiple tasks by learn- ing to mask weights

    Arun Mallya, Dillon Davis, and Svetlana Lazebnik. Piggy- back: Adapting a single network to multiple tasks by learn- ing to mask weights. InProceedings of the European con- ference on computer vision (ECCV), pages 67–82, 2018. 3

    2018

  23. [23]

    Catastrophic inter- ference in connectionist networks: The sequential learning problem

    Michael McCloskey and Neal J Cohen. Catastrophic inter- ference in connectionist networks: The sequential learning problem. InPsychology of learning and motivation, pages 109–165. Elsevier, 1989. 2

    1989

  24. [24]

    Learning to remember: A synaptic plasticity driven framework for continual learning

    Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein, Patrick Jah- nichen, and Moin Nabi. Learning to remember: A synaptic plasticity driven framework for continual learning. InPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11321–11329, 2019. 3

    2019

  25. [25]

    Continual lifelong learning with 9 neural networks: A review.Neural networks, 113:54–71,

    German I Parisi, Ronald Kemker, Jose L Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with 9 neural networks: A review.Neural networks, 113:54–71,

  26. [26]

    DreamFusion: Text-to-3D using 2D Diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988, 2022. 2

    work page internal anchor Pith review arXiv 2022

  27. [27]

    Boosting the class-incremental learning in 3d point clouds via zero-collection-cost basic shape pre- training.arXiv preprint arXiv:2504.08412, 2025

    Chao Qi, Jianqin Yin, Meng Chen, Yingchun Niu, and Yuan Sun. Boosting the class-incremental learning in 3d point clouds via zero-collection-cost basic shape pre- training.arXiv preprint arXiv:2504.08412, 2025. 3

    work page arXiv 2025

  28. [28]

    Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017

    Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space.Advances in neural information processing systems, 30, 2017. 2, 7

    2017

  29. [29]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021. 2, 4, 6, 1

    2021

  30. [30]

    icarl: Incremental classifier and representation learning

    Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. InProceedings of the IEEE con- ference on Computer Vision and Pattern Recognition, pages 2001–2010, 2017. 2, 3

    2001

  31. [31]

    Continual learning in 3d point clouds: Employing spectral techniques for exemplar selec- tion

    Hossein Resani, Behrooz Nasihatkon, and Moham- madreza Alimoradi Jazi. Continual learning in 3d point clouds: Employing spectral techniques for exemplar selec- tion. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2921–2931. IEEE, 2025. 3

    2025

  32. [32]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2

    2022

  33. [33]

    Progressive Neural Networks

    Andrei A Rusu, Neil C Rabinowitz, Guillaume Desjardins, Hubert Soyer, James Kirkpatrick, Koray Kavukcuoglu, Raz- van Pascanu, and Raia Hadsell. Progressive neural networks. arXiv preprint arXiv:1606.04671, 2016. 3

    work page internal anchor Pith review arXiv 2016

  34. [34]

    Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural information processing systems, 35:36479–36494, 2022. 2

    2022

  35. [35]

    Continual learning with deep generative replay.Advances in neural information processing systems, 30, 2017

    Hanul Shin, Jung Kwon Lee, Jaehong Kim, and Jiwon Kim. Continual learning with deep generative replay.Advances in neural information processing systems, 30, 2017. 3

    2017

  36. [36]

    S., Y .-C

    James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual dif- fusion: Continual customization of text-to-image diffusion with c-lora.arXiv preprint arXiv:2304.06027, 2023. 3

    work page arXiv 2023

  37. [37]

    Using shape to categorize: Low-shot learning with an explicit shape bias

    Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. InProceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 1798–1808, 2021. 2, 4, 1

    2021

  38. [38]

    Create your world: Lifelong text-to- image diffusion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):6454–6470, 2024

    Gan Sun, Wenqi Liang, Jiahua Dong, Jun Li, Zhengming Ding, and Yang Cong. Create your world: Lifelong text-to- image diffusion.IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(9):6454–6470, 2024. 3

    2024

  39. [39]

    Rethinking the inception archi- tecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception archi- tecture for computer vision. InProceedings of the IEEE con- ference on computer vision and pattern recognition, pages 2818–2826, 2016. 7

    2016

  40. [40]

    arXiv preprint arXiv:2309.16653 , year=

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation.arXiv preprint arXiv:2309.16653,

    work page arXiv

  41. [41]

    Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior

    Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-it-3d: High-fidelity 3d creation from a single image with diffusion prior. InPro- ceedings of the IEEE/CVF international conference on com- puter vision, pages 22819–22829, 2023. 2

    2023

  42. [42]

    The surprising positive knowledge transfer in contin- ual 3d object shape reconstruction

    Anh Thai, Stefan Stojanov, Zixuan Huang, and James M Rehg. The surprising positive knowledge transfer in contin- ual 3d object shape reconstruction. In2022 International Conference on 3D Vision (3DV), pages 209–218. IEEE,

  43. [43]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion.Advances in neural information processing systems, 36: 8406–8441, 2023

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion.Advances in neural information processing systems, 36: 8406–8441, 2023. 2

    2023

  44. [44]

    Memory replay gans: Learning to generate new categories without forgetting.Advances in neural information processing systems, 31, 2018

    Chenshen Wu, Luis Herranz, Xialei Liu, Joost Van De Wei- jer, Bogdan Raducanu, et al. Memory replay gans: Learning to generate new categories without forgetting.Advances in neural information processing systems, 31, 2018. 3

    2018

  45. [45]

    Structured 3d latents for scalable and versatile 3d gen- eration

    Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d gen- eration. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21469–21480, 2025. 2, 4, 5, 6, 1

    2025

  46. [46]

    Explicit inductive bias for transfer learning with convolutional net- works

    LI Xuhong, Yves Grandvalet, and Franck Davoine. Explicit inductive bias for transfer learning with convolutional net- works. InInternational conference on machine learning, pages 2825–2834. PMLR, 2018. 6, 1

    2018

  47. [47]

    Contin- ual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Contin- ual learning through synaptic intelligence. InInternational conference on machine learning, pages 3987–3995. PMLR,

  48. [48]

    Static-dynamic co-teaching for class-incremental 3d object detection

    Na Zhao and Gim Hee Lee. Static-dynamic co-teaching for class-incremental 3d object detection. InProceedings of the AAAI Conference on Artificial Intelligence, pages 3436– 3445, 2022. 2, 3

    2022

  49. [49]

    Sdcot++: Improved static-dynamic co- teaching for class-incremental 3d object detection.IEEE Transactions on Image Processing, 2024

    Na Zhao, Peisheng Qian, Fang Wu, Xun Xu, Xulei Yang, and Gim Hee Lee. Sdcot++: Improved static-dynamic co- teaching for class-incremental 3d object detection.IEEE Transactions on Image Processing, 2024. 2

    2024

  50. [50]

    Da-cil: Towards domain adap- tive class-incremental 3d object detection.arXiv preprint arXiv:2212.02057, 2022

    Ziyuan Zhao, Mingxi Xu, Peisheng Qian, Ramanpreet Singh Pahwa, and Richard Chang. Da-cil: Towards domain adap- tive class-incremental 3d object detection.arXiv preprint arXiv:2212.02057, 2022. 3 10 ReConText3D: Replay-based Continual Text-to-3D Generation Supplementary Material This supplementary document provides additional de- tails supporting the main ...

    work page arXiv 2022

  51. [51]

    Additional Details on Toys4K-CL Bench- mark In this section, we present the additional details regarding ourToys4K-CL Benchmark. Benchmark details.Toys4K-CL consists of 45 base and 45 novel classes selected to maximize semantic diversity while preserving the natural long-tailed distribution of the Toys4K dataset [37]. Based on an analysis of the dataset’s...

  52. [52]

    Figure 8 shows that our count-aware allocation mitigates long-tail bias by allowing smooth budget growth while preventing large classes from monopolizing memory

    Additional Details on ReConText3D Replay Creation Our ReConText3D replay strategy provides a principled balance between semantic coverage and class proportion- ality of the replayed examplers. Figure 8 shows that our count-aware allocation mitigates long-tail bias by allowing smooth budget growth while preventing large classes from monopolizing memory

  53. [53]

    Base-class performance.Table 3 reports class-wise CLIP [29] scores on base-class assets for bothTRELLIS- XL[45] (left) andShap-E[13] (right) across the base- lines

    Additional Quantitative Results In this section, we provide detailed class-wise CLIP scores for both base and novel categories on the Toys4K-CL benchmark, complementing the aggregated results in the main paper. Base-class performance.Table 3 reports class-wise CLIP [29] scores on base-class assets for bothTRELLIS- XL[45] (left) andShap-E[13] (right) acros...

  54. [54]

    A 3D low-poly bunch of vivid purple dodecahedron grapes arranged in a conical cluster, with a short brown cylindrical stem at the top

    Additional Qualitative Results In this section, we provide extended qualitative compar- isons. While CLIP scores are a useful metric for semantic alignment, qualitative inspection remains essential for as- sessing geometric fidelity, texture realism, structural consis- tency, and the extent of catastrophic forgetting in continual text-to-3D generation. Ba...

    work page arXiv 2035