Pith. sign in

REVIEW 5 major objections 7 minor 31 references

LVLM-Composer's Explicit Planning for Image Generation

T0 review · 5 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LVLM-Composer claims that an LVLM that decomposes a prompt into explicit visual planning tokens—objects, attributes, locations, and relations—and aligns generation to those tokens achieves the highest compositional text-to-image scores on…

desk verdict A sensible-sounding system idea with no implementation details and an evaluation setup that may be optimizing the test judges; desk-reject. read the letter →

arxiv 2507.04152 v1 pith:F6M2VSSK submitted 2025-07-05 cs.CV

Spencer Ramsey , Jeffrey Lee , Amina Grant This is my paper
classification cs.CV
keywords LVLM-Composercompositionaltext-to-imagegenerationvisualplanningtokenshierarchicalsemanticfine-grainedfeaturealignmentreinforcementlearningwithself-correctionLongBench-T2Imultimodal
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

This paper tries to establish that explicit visual planning, rather than implicit end-to-end learning, is what lets a text-to-image LVLM handle complex multi-object prompts. It proposes LVLM-Composer, a 10-billion-parameter model that first parses a prompt into structured planning tokens and then uses those tokens to condition image generation, trained in two stages: semantic-visual grounding pretraining and reinforcement learning with self-correction. If the claim holds, compositional weaknesses of lightweight LVLMs can be reduced by making the model articulate its visual strategy before drawing. The paper supports the claim with LongBench-T2I scores from two automatic evaluators, an ablation study, and a human preference study.

What carries the argument

The central object is the planning token, written $p_i = [p_i^{\text{object}}, p_i^{\text{attribute}}, p_i^{\text{location}}, p_i^{\text{relation}}]$, a vector that explicitly encodes one scene element's class, properties, spatial coordinates, and relationships to other elements. The Hierarchical Semantic Planning Module (HSPM) is a Transformer encoder that converts raw text features into a sequence of these tokens, and the Fine-Grained Feature Alignment Mechanism (FFAM) injects the tokens into the image generator through cross-attention and perceptual losses at multiple scales. The two-stage training first grounds predicted tokens against images annotated with boxes, masks, and relational graphs, then runs reinforcement learning in which a reward model scores generated images and a self-correction loop lets the planner revise low-scoring token sequences.

What would settle it

Re-score every generated image from Table I with a third judge that was never used in training or reward modeling, then compare the resulting rank order and pairwise preference with the reported human win rates; if the 0.07–0.12 point advantages shrink to noise or reverse, the central performance claim collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that LVLM-Composer consistently achieves the highest average scores on LongBench-T2I under both Gemini-2.0-Flash and InternVL3-78B evaluation, significantly outperforming state-of-the-art lightweight baselines across nearly every compositional dimension. Concretely, under Gemini-2.0-Flash it reports an average of 2.31 versus 2.24 for Janus-Pro-1B, with object score up by 0.12 and composition score up by 0.09; the InternVL3-78B evaluation shows a similar pattern. The authors attribute the gains to the combination of a Hierarchical Semantic Planning Module that turns text into structured visual blueprints and a Fine-Grained Feature Alignment Mechanism that keeps generated features tied to those blueprints throughout synthesis. Ablations removing either module degrade the average score by 0.11 to 0.15 points, and human evaluators preferred LVLM-Composer in 46.5% of forced-choice comparisons, which the paper presents as perceptual confirmation of the automatic scores.

Load-bearing premise

The load-bearing premise is that the automatic judges Gemini-2.0-Flash and InternVL3-78B are faithful, noise-free measures of compositional fidelity, and that the Stage 2 reward model—built from similar automated evaluations—does not simply teach LVLM-Composer to game those specific judges.

Editorial extensions

If this is right

  • The reported average advantage over Janus-Pro-1B is 0.07 to 0.12 points depending on the evaluator, concentrated in object, composition, and pose scores, so the practical gain is specifically in compositional fidelity rather than raw image quality.
  • Removing the HSPM drops the average score by 0.15 points and removing the FFAM drops it by 0.11 points, so both planning and alignment are load-bearing components of the claimed improvement.
  • The paper's complexity breakdown shows the gap over baselines widens on complex prompts, implying explicit planning helps most when scenes contain many objects and spatial relations.
  • The human study's 46.5% win rate indicates that the automatic-score advantage corresponds to a perceptible preference, not only a metric artifact.
  • Because planning tokens are generated before the image, the model's output plan could be inspected and edited, making the generation process more controllable than a single black-box pass.

Reading between the lines

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

  • The reported 0.07–0.12 point gaps are small enough that judge noise could change the ranking; re-scoring with a third independent evaluator and reporting variance would test whether the advantage is stable.
  • The Stage 2 reward model is described as trained on human preferences and high-fidelity automated evaluations, so the final evaluators may be part of the same feedback loop; comparing against a held-out judge not used in reward modeling would separate genuine compositional skill from reward hacking.
  • Planning tokens are human-readable, so an untested extension is interactive generation where a user edits the plan—moving an object or changing an attribute—and the image regenerates predictably from the edited tokens.
  • The same decomposition-before-generation recipe could transfer to video generation or instruction-based image editing, where compositional dependencies across frames and edits are equally hard, though the paper does not test those settings.
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

5 major / 7 minor

Summary. The paper proposes LVLM-Composer, a 10-billion-parameter LVLM for compositional text-to-image generation. It introduces a Hierarchical Semantic Planning Module (HSPM) that maps text to structured planning tokens (objects, attributes, locations, relations) and a Fine-Grained Feature Alignment Mechanism (FFAM) that conditions generation on these tokens. Training is multi-stage: Stage 1 supervises planning-token prediction with image-text-grounding data; Stage 2 uses reinforcement learning with a reward model and a self-correction module. Experiments on LongBench-T2I with Gemini-2.0-Flash and InternVL3-78B automatic scoring report small average-score advantages over Janus-Pro-1B, LLaVA-1.5 7B + SDXL, and DeepFloyd IF, plus ablations and a human preference study. The central claim is that the explicit planning architecture and RL self-correction yield consistently higher compositional fidelity.

Significance. If substantiated, the idea of making compositional planning explicit through structured tokens and refining it via RL self-correction would be a useful contribution to text-to-image generation, particularly for attribute binding and spatial relations. The paper uses an external benchmark and human raters, so the evaluation is not circular by construction. However, the reported evidence is far from sufficient: the automatic judges may overlap with the training reward model, no statistical validation is provided, and implementation details are almost entirely absent. The potential significance is therefore not realized in the manuscript as written.

major comments (5)
  1. [Section IV-B, Table I] The headline claim that LVLM-Composer 'consistently achieves the highest average scores' and 'significantly outperforms' baselines is not supported by the evidence. The average-score advantages over Janus-Pro-1B are 0.07 (Gemini-2.0-Flash) and 0.08 (InternVL3-78B), and per-dimension differences range from 0.00 to 0.12. No variance, confidence intervals, significance tests, or number of test prompts are reported. On a 1-5 scale, these differences are within the plausible noise range of the LLM judges, so the central claim needs statistical grounding before it can be accepted.
  2. [Section III-B2, Eqs. (9)-(11) vs. Section IV-A] Stage 2 optimizes the policy against reward R(Igen,T) computed by a reward model trained on 'human preferences and high-fidelity automated evaluations,' while Section IV-A uses Gemini-2.0-Flash and InternVL3-78B as the authoritative evaluators on the same compositional dimensions. The manuscript does not state whether these two judges were part of the RM training set or otherwise excluded. If they were included, the RL procedure directly optimizes the evaluation metric, and the Table I gains could reflect reward hacking rather than improved compositional fidelity. The authors must specify the RM provenance and report human-judge correlation and inter-judge agreement.
  3. [Section III-A and IV-A] The method section does not provide enough implementation detail to reproduce or verify the architecture. Eqs. (2)-(4) are generic transformer and cross-attention equations; the backbone LVLM is unnamed; the 10-billion parameter configuration is not specified; and the training data, dataset size, hyperparameters, and compute budget are missing. Section IV-A only says hyperparameters were 'rigorously tuned' without listing them. This is a load-bearing gap because the claimed benefits of HSPM and FFAM cannot be assessed or reproduced.
  4. [Section IV-D, Table III] The human evaluation is the only independent check on the automatic scores, but it is not statistically grounded. It uses 500 prompts and 30 raters, reports no confidence intervals, significance tests, or inter-rater agreement, and the 46.5% three-way win rate (vs. 28.5% and 25.0%; chance = 33%) is suggestive but not conclusive. The paper should provide per-item variance, raters' agreement, and a test of the win-rate differences.
  5. [Section IV-C, Table II] The ablation conclusions that HSPM and FFAM are 'indispensable' rest on the same unvalidated Gemini-2.0-Flash judge, with no error bars or significance tests (e.g., the average drops from 2.31 to 2.16 or 2.20). Moreover, the self-correction mechanism described in Eq. (12) is never ablated, so its contribution to the results is unverified.
minor comments (7)
  1. [Section III-B1, Eq. (7)] Equation (7) has mismatched parentheses and does not close the sum; please fix and define all loss terms.
  2. [Section III-B1, Eq. (8)] Equation (8) uses Generate(ZI|P_GT,T) but does not define Igen; clarify the relationship between the generated image and the reconstruction target.
  3. [References] Several related-work citations appear mismatched (e.g., [10] cites AT-GAN instead of the original GAN paper; [14] cites Diffusion-4k instead of Latent Diffusion Models; [13] cites a wrapped-phase denoising paper instead of DDPM).
  4. [Section IV-A] Section IV-A contains a typo ('such such') and should list the actual hyperparameter values.
  5. [General] The paper contains no qualitative examples; a figure with generated images illustrating compositional improvements would help.
  6. [Section IV-E, Tables IV-VI] Tables IV-VI report additional analyses but do not explain how prompt-complexity categories, attribute-binding scores, or 'epochs to 90% of max' were computed.
  7. [Abstract / Main text] The term '10-billion parameter scale' is used in the abstract, but the main text never gives the parameter count or identifies the base model.

Circularity Check

0 steps flagged · score 0.0 of 10

No formal circularity found: evaluation rests on an external benchmark, and the training-reward/evaluation-judge overlap is not established by the paper's own equations.

full rationale

The central performance claim is benchmarked on LongBench-T2I with external evaluators (Gemini-2.0-Flash and InternVL3-78B), and the method's training signal in Stage 2 is a reward model described only as trained on 'human preferences and high-fidelity automated evaluations.' The paper never states that this reward model is identical to the final evaluators or that the scored dimensions in Eq. 9 are computed by the same functions reported in Table I, so the suspected reward-hacking scenario remains a risk about validation independence, not a demonstrated by-construction identity. The ablation and human studies are also external checks rather than re-statements of the model's own loss. No load-bearing self-citation chain or renamed-known-result pattern appears; the cited benchmark and backbone references are independent artifacts. Accordingly, no circular step can be exhibited under the required standard, and the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim depends on unverified annotation labels, unvalidated automatic judges, an unspecified reward model, and a generic architectural description. Four free parameters are named but no values are given, and two invented constructs, planning tokens and the refinement module, have no external evidence.

free parameters (4)
  • lambda_1 (planning and reconstruction loss weight) = not reported
    Equation 6 balances L_planning and L_reconstruction, but the value is never given.
  • w_d (reward dimension weights) = not reported
    Equation 9 weights compositional dimensions in the reward; the paper says they are pre-determined or adaptive but does not specify them.
  • Planning token count N and token decomposition hyperparameters = not reported
    The number of planning tokens and the exact meaning of p_object, p_attribute, p_location, and p_relation are not defined, yet they shape the entire method.
  • Training hyperparameters (learning rate, batch size, epochs) = not reported
    Section IV-A says hyperparameters were rigorously tuned, but none are listed.
assumptions (4)
  • domain assumption Ground-truth planning tokens P_GT can be reliably derived from image-text pairs annotated with bounding boxes, masks, and relational graphs (Section III-B-1).
    The entire Stage 1 training depends on this annotation pipeline, which is not described or released.
  • domain assumption Gemini-2.0-Flash and InternVL3-78B produce valid, stable scores across the nine compositional dimensions (Section IV-A).
    No validation of these judges against human labels or against each other is provided.
  • ad hoc to paper The reward model RM in Stage 2 gives accurate error signals about which compositional aspect failed (Section III-B-2).
    The self-correction mechanism assumes the reward model's granular feedback is correct, but no reward model details or accuracy are given.
  • ad hoc to paper The generic Transformer encoder and cross-attention equations (Eqs. 2 to 4) instantiate the claimed hierarchical planning and fine-grained alignment behavior.
    Equations are placeholders; no concrete model architecture, tensor shapes, or training objective details are provided.
invented entities (2)
  • Planning tokens P = [p_object, p_attribute, p_location, p_relation]
    purpose: Intermediate representation said to encode objects, attributes, locations, and relations from the text prompt.
    The paper provides no token definition, no examples of learned tokens, no visualization, and no ablations isolating token content, so the representation has no independent falsifiable handle.
  • RefinementModule for self-correction
    purpose: Generates revised planning tokens P-prime after a low reward score (Equation 12).
    Equation 12 defines it abstractly; no implementation, training signal, or evaluation of the refinement step is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LVLM-Composer's Explicit Planning for Image Generation." pith.science (2026). https://pith.science/paper/F6M2VSSK

@misc{pith2026250704152,
  author       = {Pith},
  title        = {Pith review of: LVLM-Composer's Explicit Planning for Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F6M2VSSK}},
  note         = {Machine review of arXiv:2507.04152}
}
read the original abstract

The burgeoning field of generative artificial intelligence has fundamentally reshaped our approach to content creation, with Large Vision-Language Models (LVLMs) standing at its forefront. While current LVLMs have demonstrated impressive capabilities in text-to-image generation, they often falter when confronted with complex textual descriptions demanding precise compositional understanding and visual planning. This limitation particularly impacts the accurate rendering of multiple objects, their attributes, spatial relationships, and specific poses within intricate scenes, as evidenced by benchmarks like LongBench-T2I. To address these challenges, we introduce LVLM-Composer, a novel 10-billion parameter scale LVLM specifically engineered for enhanced compositional image synthesis. Our method incorporates a Hierarchical Semantic Planning Module for structured prompt decomposition and a Fine-Grained Feature Alignment Mechanism for precise visual guidance during generation. We propose a multi-stage training paradigm, featuring Hierarchical Semantic-Visual Grounding Pre-training and Compositional Planning Reinforcement Learning with Self-Correction, to instill robust compositional reasoning. Extensive experiments on the LongBench-T2I benchmark, utilizing automatic evaluation by Gemini-2.0-Flash and InternVL3-78B, demonstrate LVLM-Composer's superior performance across critical compositional dimensions including object accuracy, composition fidelity, and pose accuracy, significantly outperforming state-of-the-art baselines. An in-depth ablation study further validates the indispensable contribution of our proposed modules, while human evaluations confirm the perceptual superiority of our generated images. LVLM-Composer represents a significant step towards truly controllable and compositionally accurate open-ended text-to-image generation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 8 canonical work pages

  1. [1]

    Less is more: Vision representation compression for efficient video generation with large language models,

    Y . Zhou, J. Zhang, G. Chen, J. Shen, and Y . Cheng, “Less is more: Vision representation compression for efficient video generation with large language models,” 2024

    2024

  2. [2]

    Janus-pro: Unified multimodal understanding and generation with data and model scaling,

    X. Chen, Z. Wu, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, and C. Ruan, “Janus-pro: Unified multimodal understanding and generation with data and model scaling,” arXiv preprint arXiv:2501.17811 , 2025. 9

    arXiv 2025

  3. [3]

    Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,

    Y . Zhou, J. Yuan, and Q. Wang, “Draw all your imagine: A holistic benchmark and agent framework for complex instruction-based image generation,” arXiv preprint arXiv:2505.24787 , 2025

    arXiv 2025

  4. [4]

    Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,

    C. Wang, Y . Zhou, Q. Wang, Z. Wang, and K. Zhang, “Complexbench- edit: Benchmarking complex instruction-driven image editing via com- positional dependencies,” arXiv preprint arXiv:2506.12830 , 2025

    arXiv 2025

  5. [5]

    Weak to strong generalization for large language models with multi-capabilities,

    Y . Zhou, J. Shen, and Y . Cheng, “Weak to strong generalization for large language models with multi-capabilities,” in The Thirteenth International Conference on Learning Representations , 2025. [Online]. Available: https://openreview.net/forum?id=N1vYivuSKq

    2025

  6. [6]

    Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,

    D. Zhu, W. Shi, Z. Shi, Z. Ren, S. Wang, L. Yan, and D. Yin, “Divide-then-aggregate: An efficient tool learning method via parallel tool invocation,” arXiv preprint arXiv:2501.12432 , 2025

    arXiv 2025

  7. [7]

    Thread of thought unraveling chaotic contexts,

    Y . Zhou, X. Geng, T. Shen, C. Tao, G. Long, J.-G. Lou, and J. Shen, “Thread of thought unraveling chaotic contexts,” arXiv preprint arXiv:2311.08734, 2023

    arXiv 2023

  8. [8]

    Hierarchical reinforcement learning for handling sparse rewards in multi-goal navigation,

    J. Yan, B. Luo, and X. Xu, “Hierarchical reinforcement learning for handling sparse rewards in multi-goal navigation,” Artificial Intelligence Review, vol. 57, no. 6, p. 156, 2024

    work page 2024

Show all 31 references
  1. [9]

    Training latent variable models with auto-encoding variational bayes: A tutorial,

    Y . Zhi-Han, “Training latent variable models with auto-encoding variational bayes: A tutorial,” CoRR, vol. abs/2208.07818, 2022. [Online]. Available: https://doi.org/10.48550/arXiv.2208.07818

    2022 doi arXiv

  2. [10]

    AT-GAN: A generative attack model for adversarial transferring on generative adversarial nets,

    X. Wang, K. He, and J. E. Hopcroft, “AT-GAN: A generative attack model for adversarial transferring on generative adversarial nets,” CoRR, vol. abs/1904.07793, 2019. [Online]. Available: http: //arxiv.org/abs/1904.07793

    1904 arXiv

  3. [11]

    Progressive growing of gans for improved quality, stability, and variation,

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive growing of gans for improved quality, stability, and variation,” in 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings . Op...

    2018

  4. [12]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” in IEEE International Conference on Computer Vision, ICCV 2017, Venice, Italy, October 22-29, 2017 . IEEE Computer Society, 2017, pp. 2242–2251. [Onli...

    2017 doi

  5. [13]

    Wrapped phase denoising using denoising diffusion probabilistic models,

    S. Yang, J. Gao, J. Zhang, and C. Xu, “Wrapped phase denoising using denoising diffusion probabilistic models,” IEEE Geosci. Remote. Sens. Lett. , vol. 21, pp. 1–5, 2024. [Online]. Available: https: //doi.org/10.1109/LGRS.2024.3405000

    2024

  6. [14]

    Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion models,

    J. Zhang, Q. Huang, J. Liu, X. Guo, and D. Huang, “Diffusion-4k: Ultra-high-resolution image synthesis with latent diffusion models,” CoRR, vol. abs/2503.18352, 2025. [Online]. Available: https://doi.org/ 10.48550/arXiv.2503.18352

    2025 doi

  7. [15]

    Photorealistic text-to-image diffusion models with deep language understanding,

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans et al., “Photorealistic text-to-image diffusion models with deep language understanding,” Advances in neural information processing systems , vol. 35, ...

    2022

  8. [16]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. B. Girshick, “Masked autoencoders are scalable vision learners,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022 . IEEE, 2022, pp. 15 979–15 988. [Online]. ...

    2022

  9. [17]

    Improving cross-modal alignment for text- guided image inpainting,

    Y . Zhou and G. Long, “Improving cross-modal alignment for text- guided image inpainting,” in Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , 2023, pp. 3445–3456

    2023

  10. [18]

    Training medical large vision- language models with abnormal-aware feedback,

    Y . Zhou, L. Song, and J. Shen, “Training medical large vision- language models with abnormal-aware feedback,” arXiv preprint arXiv:2501.01377, 2025

    2025 arXiv

  11. [19]

    Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,

    J. Lu, D. Batra, D. Parikh, and S. Lee, “Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,” Advances in neural information processing systems , vol. 32, 2019

    2019

  12. [20]

    Visualbert: A simple and performant baseline for vision and language,

    L. H. Li, M. Yatskar, D. Yin, C. Hsieh, and K. Chang, “Visualbert: A simple and performant baseline for vision and language,” CoRR, vol. abs/1908.03557, 2019. [Online]. Available: http://arxiv.org/abs/1908.03557

    1908 arXiv

  13. [21]

    LXMERT: learning cross-modality encoder representations from transformers,

    H. Tan and M. Bansal, “LXMERT: learning cross-modality encoder representations from transformers,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP ...

    2019 doi

  14. [22]

    UNITER: universal image-text representation learning,

    Y . Chen, L. Li, L. Yu, A. E. Kholy, F. Ahmed, Z. Gan, Y . Cheng, and J. Liu, “UNITER: universal image-text representation learning,” in Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX , ser. Lecture Notes in Compu...

    2020 doi

  15. [23]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” in Proceedings of the 38th International Conference on Machine L...

    2021

  16. [24]

    Interleave-vla: Enhancing robot manipulation with interleaved image-text instructions,

    C. Fan, X. Jia, Y . Sun, Y . Wang, J. Wei, Z. Gong, X. Zhao, M. Tomizuka, X. Yang, J. Yan et al., “Interleave-vla: Enhancing robot manipulation with interleaved image-text instructions,” arXiv preprint arXiv:2505.02152, 2025

    2025

  17. [25]

    Sda: simple discrete augmentation for contrastive sentence representation learning,

    D. Zhu, Z. Mao, J. Lu, R. Zhao, and F. Tan, “Sda: simple discrete augmentation for contrastive sentence representation learning,” arXiv preprint arXiv:2210.03963, 2022

    2022 arXiv

  18. [26]

    Florence: A new foundation model for computer vision,

    L. Yuan, D. Chen, Y .-L. Chen, N. Codella, X. Dai, J. Gao, H. Hu, X. Huang, B. Li, C. Li et al. , “Florence: A new foundation model for computer vision,” arXiv preprint arXiv:2111.11432 , 2021

    2021 arXiv

  19. [27]

    Coca: Contrastive captioners are image-text foundation models,

    J. Yu, Z. Wang, V . Vasudevan, L. Yeung, M. Seyedhosseini, and Y . Wu, “Coca: Contrastive captioners are image-text foundation models,” Trans. Mach. Learn. Res. , vol. 2022, 2022. [Online]. Available: https://openreview.net/forum?id=Ee277P3AYC

    2022

  20. [28]

    Tx-llava: Large language and vision assistant for temporal changes in chest x-rays,

    H. Elgendy and H. Cholakkal, “Tx-llava: Large language and vision assistant for temporal changes in chest x-rays,” in 22nd IEEE International Symposium on Biomedical Imaging, ISBI 2025, Houston, TX, USA, April 14-17, 2025 . IEEE, 2025, pp. 1–4. [Online]. Available: https://doi...

    2025

  21. [29]

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

    D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “Minigpt-4: Enhancing vision-language understanding with advanced large language models,” arXiv preprint arXiv:2304.10592 , 2023

    2023 arXiv

  22. [30]

    Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,

    D. Zhu, X. Tang, W. Han, J. Lu, Y . Zhao, G. Xing, J. Wang, and D. Yin, “Vislinginstruct: Elevating zero-shot learning in multi-modal lan- guage models with autonomous instruction optimization,” arXiv preprint arXiv:2402.07398, 2024

    2024 arXiv

  23. [31]

    Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,

    Y . Zhou, L. Song, and J. Shen, “Mam: Modular multi-agent framework for multi-modal medical diagnosis via role-specialized collaboration,” arXiv preprint arXiv:2506.19835 , 2025. 10

    2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.