Pith. sign in

REVIEW 3 major objections 4 minor 80 references

T2I-ConBench: Text-to-Image Benchmark for Continual Post-training

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper introduces a unified benchmark for continual post-training of text-to-image diffusion models, measuring not only task performance but also retention of pretrained abilities, forgetting, and cross-task generalization…

desk verdict T2I-ConBench is a genuinely useful benchmark artifact, but its headline comparative claims rest on an unvalidated VQA pipeline, and one table row looks like a transcription error. read the letter →

arxiv 2505.16875 v1 pith:VNNLKJUL submitted 2025-05-22 cs.CV cs.AIcs.LG

Zhehao Huang , Yuhang Liu , Yixin Lou , Zhengbao He , Mingzhen He , Wenxing Zhou , Tao Li , Kehan Li
show 2 more authors
Zeyi Huang Xiaolin Huang
This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords text-to-imagediffusioncontinualpost-trainingcatastrophicforgettingcross-taskgeneralizationitemcustomizationdomainenhancementbenchmarkvision-languageQA
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Continual post-training—sequentially adapting a single text-to-image diffusion model to new tasks—is usually evaluated only by how well the model does the latest task. This paper argues that such an evaluation misses the trade-offs that actually matter: how much the model forgets pretrained concepts, how much earlier-task performance decays, and whether knowledge from separate tasks can be recombined into novel compositions. To make those trade-offs measurable, T2I-ConBench fixes the base model and datasets, curates two types of tasks (item customization and domain enhancement), and scores every method along four axes: preservation of generality, target-task performance, catastrophic forgetting, and cross-task generalization. Benchmarking ten representative methods over three task sequences, the paper reports that no method dominates, that joint 'oracle' training is not an upper bound in imbalanced streams, and that cross-task generalization is an unsolved challenge. If correct, the benchmark supplies a common protocol for comparing stability and plasticity in image-model post-training rather than isolated accuracy numbers.

What carries the argument

The load-bearing machinery is the automated evaluation pipeline and the curated task streams. The pipeline combines FID on MS-COCO for general generation quality, a complex-composition subset of T2I-CompBench for text-image alignment, HPS for domain aesthetics, and a vision-language question-answering pipeline that decomposes prompts into yes/no questions and scores responses. The cross-task test sets (Item+Item, Item+Domain, Domain+Domain) turn compositional generation into a measurable accuracy. The task streams—four personalized items in order, two enhanced domains in order, and mixed item-domain orders—isolate the effect of the continual-learning algorithm because the base model and datasets are held fixed.

What would settle it

Have human annotators answer the same yes/no questions on a random sample of the generated and reference images used for Unique-Sim, Class-Sim, and cross-task scores, then compare human answers with the VQA answers; near-chance agreement on a meaningful subset would show that the benchmark's item-similarity and cross-task conclusions are not established.

Watch

Extended reading notes

Core claim

With one base model and fixed task data, T2I-ConBench finds that the algorithm and task order, not just data quality, determine what a continually post-trained text-to-image model retains. On sequential item customization, most methods improve FID but LoRA variants fail to acquire later items; on domain enhancement, most methods degrade global FID while improving domain alignment; and under mixed item-domain streams, whichever task is learned second usually dominates, with replay and the joint 'oracle' both failing when the data stream is imbalanced. The paper's stated findings are that no single method wins across all four dimensions, that treating joint training as the oracle upper bound is unreliable, and that cross-task compositional generation remains an open problem even for methods that preserve individual tasks well.

Load-bearing premise

The load-bearing premise is that the automated yes/no vision-language questions measure item similarity and cross-task composition accurately, without any human calibration; if that scoring is noisy or biased, the benchmark's rankings on those axes do not follow.

Editorial extensions

If this is right

  • A method's target-task score alone is misleading; the benchmark's four dimensions show that gains on a new task can coincide with sharp drops in pretrained FID, item similarity, or cross-task mixing.
  • The 'oracle' joint-training baseline should not be treated as an unconditional upper bound, especially when tasks are imbalanced in data size and granularity.
  • Rehearsal with a simple buffer is competitive under balanced streams but can collapse on imbalanced streams, so buffer design is a first-order variable in continual post-training.
  • Parameter-isolation methods such as LoRA minimize forgetting but can fail to acquire new concepts, meaning stability and plasticity are not well separated in current approaches.
  • Cross-task generalization needs to be evaluated as a distinct capability, since preserving prior knowledge and recombining it across tasks do not go together for most baselines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable extension the paper does not pursue: replacing the single vision-language evaluator with a panel of VLMs and a small human-annotated subset would show whether the reported rankings of methods on Unique-Sim and cross-task accuracy are stable or an artefact of one evaluator.
  • The paper's result that task order changes which knowledge survives suggests an implicit design principle—schedule coarse domains before fine-grained items—that could be tested as an explicit curriculum rule in future work.
  • Because all conclusions about item similarity and composition ride on binary VQA answers, publishing per-prompt failures would let readers distinguish genuine generative limits from evaluator limits; this is an available follow-up study, not a claim of the paper.
  • The benchmark's structure could be adapted as a regression suite for any new post-training method, since it isolates the algorithm while holding architecture and data fixed; the authors release the code, but the standalone value of that protocol is an editorial extrapolation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces T2I-ConBench, a benchmark for continual post-training of text-to-image diffusion models. It defines two task scenarios (item customization and domain enhancement), four evaluation dimensions (pretrained capability preservation, downstream performance, catastrophic forgetting, and cross-task generalization), and an automated evaluation pipeline combining FID, T2I-CompBench, Human Preference Score, and a VQA-based pipeline built on Qwen2.5-7B-Instruct. The authors evaluate ten continual learning baselines plus a joint-training oracle on three task sequences across two base models (PixArt-α and SD v1.4), reporting that no method excels on all fronts, that joint training is not a panacea, and that cross-task generalization remains an open challenge.

Significance. If the evaluation pipeline is reliable, T2I-ConBench is a valuable community asset: it fixes the base model and task datasets, covers two practically important post-training granularities, adds a cross-task compositional dimension that prior benchmarks lack, and the authors release datasets, code, and evaluation tools. The paper is also appropriately candid in its limitations section (Appendix B) about synthetic-data bias and architecture scope. However, the benchmark's validity as a standardized protocol currently hinges on an unvalidated automated VQA pipeline and single-run point estimates; the headline comparative findings are suggestive rather than established. With human calibration, error bars, and a few table fixes, the paper would make a solid contribution to evaluation methodology for continual T2I post-training.

major comments (3)
  1. [Appendix E, Eqs. (A4)-(A5); Tables 1-2] Every item-similarity, class-similarity, and cross-task metric in the paper reduces to binary yes/no answers from Qwen2.5-7B-Instruct, yet the manuscript reports no human validation, no calibration study, and no confidence intervals for this VQA pipeline. The headline cross-task comparisons involve small margins (e.g., Domain+Domain 0.4279 for MoFO versus 0.4577 for Joint and 0.4179 for Replay in Table 1; Item+Item 0.3694 for Replay versus 0.3500 for SeqLoRA), so a systematic bias of a few percentage points on rare species (Spix's macaw, gerenuk) or personalized items could flip the rankings and the 'cross-task generalization remains an open challenge' takeaway. The authors should calibrate the VQA answers against human judgments on a held-out subset and report agreement statistics and confidence intervals, or at minimum demonstrate stability across multiple VLMs and repeated sampling.
  2. [Section 6.1, Tables 1-2] All results in Tables 1, 2, and A4 are reported as single point estimates without multiple seeds or error bars. The central empirical claims—'no single method excels everywhere,' 'Joint is not a panacea,' and 'Replay performs well under balanced streams'—are comparisons of point estimates that are often within a few percentage points (e.g., Comp scores around 0.34-0.35 across nearly all methods in Table 1). Without repeated runs and a measure of variance, the claim that one method outperforms another is not statistically supported. Please add seed variation and report means and standard deviations, or at least state that differences are qualitative.
  3. [Table 2, Order 1 (HFT and MoFO rows)] In Table 2 Order 1, the HFT and MoFO rows are identical across eight of the nine reported metrics (FID 28.8221, Comp 0.3500, Body-HPS 0.3020, Nature-HPS 0.2827, I+I 0.2444, I+D 0.3918, D+D 0.3930, Class-Sim 0.0300), differing only in Unique-Sim (0.0375 vs 0.0350). This is almost certainly a copy/transcription error, since the two methods have distinct update rules and their results differ elsewhere both in Table 1 and in Table A4. The authors must verify the raw logs and correct the table; if the numbers are genuinely identical, an explanation is required.
minor comments (4)
  1. [Section 1, paragraph 2] The phrase 'T2I suffercatastrophic forgetting' is missing a space and should read 'T2I suffer catastrophic forgetting'.
  2. [Appendix E, first paragraph] The text 'text graph model' appears twice and should presumably be 'text-to-image model' or 'text-to-image diffusion model'.
  3. [References] Reference [35] is incomplete; it should provide the full citation for FLUX rather than just 'Black Forest Labs. Flux. 2024.'
  4. [Appendix E, Table A2] In the 'Class Unique Question Template' section, the wording for the sneaker row uses 'style' while other rows use 'breed'; please clarify whether 'style' is intentional and consistent with the intended similarity question.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: T2I-ConBench is a measurement instrument whose empirical takeaways are summaries over externally defined metrics, not derivations from their own outputs.

full rationale

The paper is a benchmark and empirical evaluation contribution, not a derivation chain. Its central claims—no single method excels everywhere, joint 'oracle' training is not a panacea, and cross-task generalization remains unsolved—are observational summaries over a fixed set of evaluation metrics: FID, T2I-CompBench, HPS, and the VQA-based Unique-Sim, Class-Sim, and cross-task accuracies (Eqs. A1–A5). None of these metrics is defined in terms of the paper's conclusions, and no parameter is fitted to a subset of data and then reported as a prediction of a closely related quantity. The baselines (SeqFT, Replay, EWC, ℓ2-norm, HFT, MoFO, SeqLoRA, IncLoRA, O-LoRA, C-LoRA) are standard external methods with cited origins, not self-cited load-bearing premises, and no uniqueness theorem or ansatz is imported from the authors' own prior work. The main epistemic risk is the unvalidated Qwen2.5-7B-Instruct VQA pipeline used for item-similarity and cross-task scores; the paper itself acknowledges that its automated metrics may carry biases (Appendix B, Limitations). However, an uncalibrated or unvalidated measurement instrument is a validity and reproducibility concern, not circularity: the benchmark's conclusions do not reduce by construction to the metric definitions, and the empirical rankings could in principle be wrong, which is exactly what makes the claims non-tautological. The suspicious near-identical HFT/MoFO rows in Table 2 Order 1 are a possible transcription or release-check issue, not a circular-reasoning pattern. Overall, no step in the claimed contribution is equivalent to its own inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The benchmark draws on standard metrics and assumptions about their validity, detailed in Section 4 and Appendix E. No free parameters are fitted to data; all choices are design decisions. No new entities are introduced.

assumptions (5)
  • domain assumption Standard T2I metrics (FID, T2I-CompBench, HPS) and VQA scores are valid proxies for the four evaluated dimensions (generality, target performance, forgetting, cross-task generalization).
    Used throughout Section 4 and Appendix E; no calibration against human ratings is reported.
  • domain assumption The Qwen2.5-7B-Instruct VLM answers accurately reflect image-text semantic alignment for item similarity and compositional prompts.
    Appendix E 'Answer Scoring' uses binary yes/no from the VLM without human verification.
  • domain assumption Synthetic images generated by Flux_dev and manually filtered are suitable training data for domain enhancement.
    Section 3 describes generation and manual screening; manual filtering is subjective and not fully reproducible.
  • domain assumption PixArt-alpha and Stable Diffusion v1.4 are representative base models, so conclusions transfer to other T2I models.
    Section 6.1 and Appendix H use these two architectures; the limitation section acknowledges mid-scale only.
  • domain assumption Joint training is a meaningful oracle upper bound for continual post-training.
    Section 5 defines Joint as oracle, but Table 2 shows it underperforms on imbalanced streams, so this assumption is contested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of T2I-ConBench: Text-to-Image Benchmark for Continual Post-training." pith.science (2026). https://pith.science/paper/VNNLKJUL

@misc{pith2026250516875,
  author       = {Pith},
  title        = {Pith review of: T2I-ConBench: Text-to-Image Benchmark for Continual Post-training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VNNLKJUL}},
  note         = {Machine review of arXiv:2505.16875}
}
read the original abstract

Continual post-training adapts a single text-to-image diffusion model to learn new tasks without incurring the cost of separate models, but naive post-training causes forgetting of pretrained knowledge and undermines zero-shot compositionality. We observe that the absence of a standardized evaluation protocol hampers related research for continual post-training. To address this, we introduce T2I-ConBench, a unified benchmark for continual post-training of text-to-image models. T2I-ConBench focuses on two practical scenarios, item customization and domain enhancement, and analyzes four dimensions: (1) retention of generality, (2) target-task performance, (3) catastrophic forgetting, and (4) cross-task generalization. It combines automated metrics, human-preference modeling, and vision-language QA for comprehensive assessment. We benchmark ten representative methods across three realistic task sequences and find that no approach excels on all fronts. Even joint "oracle" training does not succeed for every task, and cross-task generalization remains unsolved. We release all datasets, code, and evaluation tools to accelerate research in continual post-training for text-to-image models.

Figures

Figures reproduced from arXiv: 2505.16875 by the authors.

Figure 1
Figure 1. Overview of T2I-ConBench. Our benchmark consists of four components: (1) challenging [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Body pose distribution. For Domain Enhancement tasks, we specifically fo￾cus on two domains: natural world concepts and human portraits, which we refer to as “Nature” and “Body” domains, respectively. To enhance the base model’s im￾age generation quality and semantic alignment within these domains, we first generate numerous prompts con￾taining various concepts within each domain. We then use the base model to test … view at source ↗
Figure 3
Figure 3. Evaluation pipeline of cross-task generalization. Cross-task Generalization We generate prompts that merge concepts from differ￾ent tasks and assess whether the fine-tuned model can accurately render these novel combinations. We also score cross-task per￾formance using a VQA pipeline ( [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Overview of the continual post-training baselines evaluated in this work, encompass￾ing rehearsal-based, regularization-based, and parameter-isolation methods (sparse fine-tuning and low-rank adaptation). These baselines are described in Sec. 5 and Appendix F. Regulari…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

80 extracted references · 59 canonical work pages

  1. [1]

    Denton, Seyed Kam- yar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kam- yar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding.NeurIPS, 2022

    work page 2022

  2. [2]

    High-resolution image synthesis with latent diffusion models.CVPR, 2022

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models.CVPR, 2022

    work page 2022

  3. [3]

    SDXL: improving latent diffusion models for high-resolution image synthesis.ICLR, 2024

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: improving latent diffusion models for high-resolution image synthesis.ICLR, 2024

    work page 2024

  4. [4]

    Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li

    Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- α: Fast training of diffusion transformer for photorealistic text-to-image synthesis.ICLR, 2024

    work page 2024

  5. [5]

    Pixart- Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation.ECCV, 2024

    Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- Σ: Weak-to-strong training of diffusion transformer for 4k text-to-image generation.ECCV, 2024

    work page 2024

  6. [6]

    A survey on personalized content synthesis with diffusion models.arXiv preprint arXiv:2405.05538, 2025

    Xulu Zhang, Xiaoyong Wei, Wentao Hu, Jinlin Wu, Jiaxin Wu, Wengyu Zhang, Zhaoxiang Zhang, Zhen Lei, and Qing Li. A survey on personalized content synthesis with diffusion models.arXiv preprint arXiv:2405.05538, 2025

    arXiv 2025

  7. [7]

    Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation.CVPR, 2023

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation.CVPR, 2023

    work page 2023

  8. [8]

    Prompt to polyp: Clinically- aware medical image synthesis with diffusion models.arXiv preprint arXiv:2505.05573, 2025

    Mikhail Chaichuk, Sushant Gautam, Steven Hicks, and Elena Tutubalina. Prompt to polyp: Clinically- aware medical image synthesis with diffusion models.arXiv preprint arXiv:2505.05573, 2025

    arXiv 2025

Show all 80 references
  1. [9]

    Customizing text-to-image diffusion with object viewpoint control.SIGGRAPH Asia, 2024

    Nupur Kumari, Grace Su, Richard Zhang, Taesung Park, Eli Shechtman, and Jun-Yan Zhu. Customizing text-to-image diffusion with object viewpoint control.SIGGRAPH Asia, 2024

    2024

  2. [10]

    Diffusion based augmenta- tion for captioning and retrieval in cultural heritage.ICCV (Workshops), 2023

    Dario Cioni, Lorenzo Berlincioni, Federico Becattini, and Alberto Del Bimbo. Diffusion based augmenta- tion for captioning and retrieval in cultural heritage.ICCV (Workshops), 2023

    2023

  3. [11]

    Parameter-efficient transfer learning for NLP.ICML, 2019

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP.ICML, 2019

    2019

  4. [12]

    Jonathan Pilault, Amine Elhattami, and Christopher J. Pal. Conditionally adaptive multi-task learning: Improving transfer learning in NLP using fewer parameters & less data.ICLR, 2021

    2021

  5. [13]

    Towards a unified view of parameter-efficient transfer learning.ICLR, 2022

    Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning.ICLR, 2022

    2022

  6. [14]

    Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities.NPJ Computational Materials, 2025

    Wei Lu, Rachel K Luu, and Markus J Buehler. Fine-tuning large language models for domain adaptation: Exploration of training strategies, scaling, model merging and synergistic capabilities.NPJ Computational Materials, 2025

    2025

  7. [15]

    Continual diffusion: Continual customization of text-to-image diffusion with c-lora.Trans

    James Seale Smith, Yen-Chang Hsu, Lingyu Zhang, Ting Hua, Zsolt Kira, Yilin Shen, and Hongxia Jin. Continual diffusion: Continual customization of text-to-image diffusion with c-lora.Trans. Mach. Learn. Res., 2024

    2024

  8. [16]

    Continual training of language models for few-shot learning.EMNLP, 2022

    Zixuan Ke, Haowei Lin, Yijia Shao, Hu Xu, Lei Shu, and Bing Liu. Continual training of language models for few-shot learning.EMNLP, 2022

    2022

  9. [17]

    Continual pre-training of language models.ICLR, 2023

    Zixuan Ke, Yijia Shao, Haowei Lin, Tatsuya Konishi, Gyuhak Kim, and Bing Liu. Continual pre-training of language models.ICLR, 2023

    2023

  10. [18]

    Robert M. French. Catastrophic interference in connectionist networks: Can it be predicted, can it be prevented?NIPS, 1993

    1993

  11. [19]

    Connectionist models of recognition memory: constraints imposed by learning and forgetting functions.Psychological Review, 1990

    Roger Ratcliff. Connectionist models of recognition memory: constraints imposed by learning and forgetting functions.Psychological Review, 1990

    1990

  12. [20]

    A comprehensive survey of continual learning: Theory, method and application.IEEE Trans

    Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: Theory, method and application.IEEE Trans. Pattern Anal. Mach. Intell., 2024. 10

    2024

  13. [21]

    Dokania, Philip H

    Arslan Chaudhry, Marcus Rohrbach, Mohamed Elhoseiny, Thalaiyasingam Ajanthan, Puneet K. Dokania, Philip H. S. Torr, and Marc’Aurelio Ranzato. On tiny episodic memories in continual learning.arXiv preprint arXiv:1902.10486, 2019

    1902 arXiv

  14. [22]

    Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell

    James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting i...

    2017

  15. [23]

    Continual learning through synaptic intelligence

    Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. ICML, 2017

    2017

  16. [24]

    HFT: half fine-tuning for large language models.arXiv preprint arXiv:2404.18466, 2024

    Tingfeng Hui, Zhenyu Zhang, Shuohuan Wang, Weiran Xu, Yu Sun, and Hua Wu. HFT: half fine-tuning for large language models.arXiv preprint arXiv:2404.18466, 2024

    2024 arXiv

  17. [25]

    Mofo: Momentum-filtered optimizer for mitigating forgetting in llm fine-tuning.arXiv preprint arXiv:2407.20999, 2025

    Yupeng Chen, Senmiao Wang, Yushun Zhang, Zhihang Lin, Haozhe Zhang, Weijian Sun, Tian Ding, and Ruoyu Sun. Mofo: Momentum-filtered optimizer for mitigating forgetting in llm fine-tuning.arXiv preprint arXiv:2407.20999, 2025

    2025

  18. [26]

    Continual learning of large language models: A comprehensive survey.arXiv preprint arXiv:2404.16789, 2024

    Haizhou Shi, Zihao Xu, Hengyi Wang, Weiyi Qin, Wenyuan Wang, Yibin Wang, Zifeng Wang, Sayna Ebrahimi, and Hao Wang. Continual learning of large language models: A comprehensive survey.arXiv preprint arXiv:2404.16789, 2024

    2024 arXiv

  19. [27]

    Trace: A comprehensive benchmark for continual learning in large language models.arXiv preprint arXiv:2310.06762, 2023

    Xiao Wang, Yuansen Zhang, Tianze Chen, Songyang Gao, Senjie Jin, Xianjun Yang, Zhiheng Xi, Rui Zheng, Yicheng Zou, Tao Gui, Qi Zhang, and Xuanjing Huang. Trace: A comprehensive benchmark for continual learning in large language models.arXiv preprint arXiv:2310.06762, 2023

    arXiv

  20. [28]

    Dick, and Hidenori Tanaka

    Maya Okawa, Ekdeep Singh Lubana, Robert P. Dick, and Hidenori Tanaka. Compositional abilities emerge multiplicatively: Exploring diffusion models on a synthetic task.NeurIPS, 2023

    2023

  21. [29]

    Are transformers able to reason by connecting separated knowledge in training data?arXiv preprint arXiv:2501.15857, 2025

    Yutong Yin and Zhaoran Wang. Are transformers able to reason by connecting separated knowledge in training data?arXiv preprint arXiv:2501.15857, 2025

    2025

  22. [30]

    Domainstudio: Fine-tuning diffusion models for domain-driven image generation using limited data.arXiv preprint arXiv:2306.14153, 2024

    Jingyuan Zhu, Huimin Ma, Jiansheng Chen, and Jian Yuan. Domainstudio: Fine-tuning diffusion models for domain-driven image generation using limited data.arXiv preprint arXiv:2306.14153, 2024

    2024 arXiv

  23. [31]

    Diffusion models beat gans on image synthesis.NeurIPS, 2021

    Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis.NeurIPS, 2021

    2021

  24. [32]

    T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation.NeurIPS, 2023

    Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation.NeurIPS, 2023

    2023

  25. [33]

    Geneval: An object-focused framework for evaluating text-to-image alignment.NeurIPS, 2023

    Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment.NeurIPS, 2023

    2023

  26. [34]

    Dreambench++: A human-aligned benchmark for personalized image generation.ICLR, 2025

    Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation.ICLR, 2025

    2025

  27. [35]

    Black Forest Labs. Flux. 2024

    2024

  28. [36]

    Synthetic data protection: Towards a paradigm change in data regulation?Big Data Soc., 2024

    Ana Beduschi. Synthetic data protection: Towards a paradigm change in data regulation?Big Data Soc., 2024

    2024

  29. [37]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium.NIPS, 2017

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium.NIPS, 2017

    2017

  30. [38]

    Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context.ECCV, 2014

    2014

  31. [39]

    A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges.arXiv preprint arXiv:2501.02189, 2025

    Zongxia Li, Xiyang Wu, Hongyang Du, Fuxiao Liu, Huy Nghiem, and Guangyao Shi. A survey of state of the art large vision language models: Alignment, benchmark, evaluations and challenges.arXiv preprint arXiv:2501.02189, 2025

    2025 arXiv

  32. [40]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models.ICLR, 2024

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models.ICLR, 2024

    2024

  33. [41]

    Robust visual question answering: Datasets, methods, and future challenges.IEEE Trans

    Jie Ma, Pinghui Wang, Dechen Kong, Zewei Wang, Jun Liu, Hongbin Pei, and Junzhou Zhao. Robust visual question answering: Datasets, methods, and future challenges.IEEE Trans. Pattern Anal. Mach. Intell., 2023. 11

    2023

  34. [42]

    Human preference score: Better aligning text-to-image models with human preference.ICCV, 2023

    Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference.ICCV, 2023

    2023

  35. [43]

    A theory for knowledge transfer in continual learning.CoLLAs, 2022

    Diana Benavides Prado and Patricia Riddle. A theory for knowledge transfer in continual learning.CoLLAs, 2022

    2022

  36. [44]

    Is multi-task learning an upper bound for continual learning?ICASSP, 2023

    Zihao Wu, Huy Tran, Hamed Pirsiavash, and Soheil Kolouri. Is multi-task learning an upper bound for continual learning?ICASSP, 2023

    2023

  37. [45]

    Slca++: Unleash the power of sequential fine-tuning for continual learning with pre-training.arXiv preprint arXiv:2408.08295, 2024

    Gengwei Zhang, Liyuan Wang, Guoliang Kang, Ling Chen, and Yunchao Wei. Slca++: Unleash the power of sequential fine-tuning for continual learning with pre-training.arXiv preprint arXiv:2408.08295, 2024

    2024 arXiv

  38. [46]

    Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset.arXiv preprint arXiv:2505.09568, 2025

    Jiuhai Chen, Zhiyang Xu, Xichen Pan, Yushi Hu, Can Qin, Tom Goldstein, Lifu Huang, Tianyi Zhou, Saining Xie, Silvio Savarese, Le Xue, Caiming Xiong, and Ran Xu. Blip3-o: A family of fully open unified multimodal models-architecture, training and dataset.arXiv preprint arXiv:25...

    2025 arXiv

  39. [47]

    A statistical theory of regularization-based continual learning.ICML, 2024

    Xuyang Zhao, Huiyuan Wang, Weiran Huang, and Wei Lin. A statistical theory of regularization-based continual learning.ICML, 2024

    2024

  40. [48]

    Approximate fisher information matrix to characterise the training of deep neural networks.IEEE Trans

    Zhibin Liao, Tom Drummond, Ian Reid, and Gustavo Carneiro. Approximate fisher information matrix to characterise the training of deep neural networks.IEEE Trans. Pattern Anal. Mach. Intell., 2018

    2018

  41. [49]

    BERT: pre-training of deep bidirectional transformers for language understanding.NAACL-HLT, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: pre-training of deep bidirectional transformers for language understanding.NAACL-HLT, 2019

    2019

  42. [50]

    Orthogonal subspace learning for language model continual learning.EMNLP, 2023

    Xiao Wang, Tianze Chen, Qiming Ge, Han Xia, Rong Bao, Rui Zheng, Qi Zhang, Tao Gui, and Xuanjing Huang. Orthogonal subspace learning for language model continual learning.EMNLP, 2023

    2023

  43. [51]

    Sutherland

    Yi Ren and Danica J. Sutherland. Learning dynamics of llm finetuning.ICLR, 2025

    2025

  44. [52]

    Tsvd: Bridging theory and practice in continual learning with pre-trained models.arXiv preprint arXiv:2410.00645, 2025

    Liangzu Peng, Juan Elenter, Joshua Agterberg, Alejandro Ribeiro, and René Vidal. Tsvd: Bridging theory and practice in continual learning with pre-trained models.arXiv preprint arXiv:2410.00645, 2025

    2025 arXiv

  45. [53]

    Create your world: Lifelong text-to-image diffusion.IEEE Trans

    Gan Sun, Wenqi Liang, Jiahua Dong, Jun Li, Zhengming Ding, and Yang Cong. Create your world: Lifelong text-to-image diffusion.IEEE Trans. Pattern Anal. Mach. Intell., 2024

    2024

  46. [54]

    Progressive compositionality in text-to- image generative models.arXiv preprint arXiv:2410.16719, 2025

    Evans Xu Han, Linghao Jin, Xiaofeng Liu, and Paul Pu Liang. Progressive compositionality in text-to- image generative models.arXiv preprint arXiv:2410.16719, 2025

    2025 arXiv

  47. [55]

    Emu: Enhancing image generation models using photogenic needles in a haystack.arXiv preprint arXiv:2309.15807, 2023

    Xiaoliang Dai, Ji Hou, Chih-Yao Ma, Sam Tsai, Jialiang Wang, Rui Wang, Peizhao Zhang, Simon Vandenhende, Xiaofang Wang, Abhimanyu Dubey, Matthew Yu, Abhishek Kadian, Filip Radenovic, Dhruv Mahajan, Kunpeng Li, Yue Zhao, Vladan Petrovic, Mitesh Kumar Singh, Simran Motwani, Yi W...

    2023 arXiv

  48. [56]

    U-net: Convolutional networks for biomedical image segmentation.MICCAI, 2015

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation.MICCAI, 2015

    2015

  49. [57]

    Learning transferable visual models from natural language supervision.ICML, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision.ICML, 2021

    2021

  50. [58]

    Improving image captioning with better use of captions

    Zhan Shi, Xu Zhou, Xipeng Qiu, and Xiaodan Zhu. Improving image captioning with better use of captions. Comput. Res. Repos., 2020

    2020

  51. [59]

    Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic-domain alignment.arXiv preprint arXiv:2406.04295, 2024

    Jiayi Guo, Junhao Zhao, Chaoqun Du, Yulin Wang, Chunjiang Ge, Zanlin Ni, Shiji Song, Humphrey Shi, and Gao Huang. Everything to the synthetic: Diffusion-driven test-time adaptation via synthetic-domain alignment.arXiv preprint arXiv:2406.04295, 2024

    2024 arXiv

  52. [60]

    Yu-Chuan Su, Kelvin C. K. Chan, Yandong Li, Yang Zhao, Han Zhang, Boqing Gong, Huisheng Wang, and Xuhui Jia. Identity encoder for personalized diffusion.arXiv preprint arXiv:2304.07429, 2023

    2023 arXiv

  53. [61]

    Prompt-to- prompt image editing with cross attention control.ICLR, 2022

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to- prompt image editing with cross attention control.ICLR, 2022

    2022

  54. [62]

    Diffuse-uda: Addressing unsupervised domain adaptation in medical image segmentation with appearance and structure aligned diffusion models.arXiv preprint arXiv:2408.05985, 2024

    Haifan Gong, Yitao Wang, Yihan Wang, Jiashun Xiao, Xiang Wan, and Haofeng Li. Diffuse-uda: Addressing unsupervised domain adaptation in medical image segmentation with appearance and structure aligned diffusion models.arXiv preprint arXiv:2408.05985, 2024. 12

    2024 arXiv

  55. [63]

    Diffboost: Enhancing medical image segmentation via text-guided diffusion model.IEEE Transactions on Medical Imaging, 2024

    Zheyuan Zhang, Lanhong Yao, Bin Wang, Debesh Jha, Gorkem Durak, Elif Keles, Alpay Medetalibeyoglu, and Ulas Bagci. Diffboost: Enhancing medical image segmentation via text-guided diffusion model.IEEE Transactions on Medical Imaging, 2024

    2024

  56. [64]

    Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis.CVPR, 2024

    Yuming Gu, You Xie, Hongyi Xu, Guoxian Song, Yichun Shi, Di Chang, Jing Yang, and Linjie Luo. Diffportrait3d: Controllable diffusion for zero-shot portrait view synthesis.CVPR, 2024

    2024

  57. [65]

    Towards high-fidelity 3d portrait generation with rich details by cross-view prior-aware diffusion.arXiv preprint arXiv:2411.10369, 2024

    Haoran Wei, Wencheng Han, Xingping Dong, and Jianbing Shen. Towards high-fidelity 3d portrait generation with rich details by cross-view prior-aware diffusion.arXiv preprint arXiv:2411.10369, 2024

    2024 arXiv

  58. [66]

    A note on the inception score.arXiv preprint arXiv:1801.01973, 2018

    Shane Barratt and Rishi Sharma. A note on the inception score.arXiv preprint arXiv:1801.01973, 2018

    2018 arXiv

  59. [67]

    Rethinking the inception architecture for computer vision.CVPR, 2016

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jonathon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision.CVPR, 2016

    2016

  60. [68]

    Efros, Eli Shechtman, and Oliver Wang

    Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric.CVPR, 2018

    2018

  61. [69]

    Emerging properties in self-supervised vision transformers.ICCV, 2021

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers.ICCV, 2021

    2021

  62. [70]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation.ICML, 2022

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation.ICML, 2022

    2022

  63. [71]

    Imagereward: Learning and evaluating human preferences for text-to-image generation.NeurIPS, 2023

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation.NeurIPS, 2023

    2023

  64. [72]

    Clog: Benchmarking continual learning of image generation models.arXiv preprint arXiv:2406.04584, 2024

    Haotian Zhang, Junting Zhou, Haowei Lin, Hang Ye, Jianhua Zhu, Zihao Wang, Liangcai Gao, Yizhou Wang, and Yitao Liang. Clog: Benchmarking continual learning of image generation models.arXiv preprint arXiv:2406.04584, 2024

    2024 arXiv

  65. [73]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2025

    DeepSeek-AI. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2025

    2025 arXiv

  66. [74]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  67. [75]

    Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2025

    Qwen. Qwen2.5 technical report.arXiv preprint arXiv:2412.15115, 2025

    2025 arXiv

  68. [76]

    The dawn of lmms: Preliminary explorations with gpt-4v(ision).arXiv preprint arXiv:2309.17421, 2023

    Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of lmms: Preliminary explorations with gpt-4v(ision).arXiv preprint arXiv:2309.17421, 2023

    2023 arXiv

  69. [77]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.ICLR, 2015

    2015

  70. [78]

    Zero: Memory optimizations toward training trillion parameter models.arXiv preprint arXiv:1910.02054, 2020

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models.arXiv preprint arXiv:1910.02054, 2020

    1910 arXiv

  71. [79]

    Decoupled weight decay regularization.ICLR, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.ICLR, 2019

    2019

  72. [80]

    item customization

    William Peebles and Saining Xie. Scalable diffusion models with transformers.ICCV, 2023. 13 Appendix A Related Work A.1 Large-scale Text-to-image Generative Model Large-scale text-to-image (T2I) diffusion models have rapidly become the backbone of generative AI. Building on la...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.