Pith. sign in

REVIEW 4 major objections 9 minor 38 references

Iterative Self-Improvement of Vision Language Models for Image Scoring and Self-Explanation

T0 review · 4 major / 9 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 Iterative self-training with score-conditioned preference pairs lets a vision-language model improve both its image scoring and its natural-language justifications using only the model's own generated text and an image-score dataset.

desk verdict A genuinely new self-training recipe for score-conditioned VLMs with plausible results, but the explanation-quality claims rest on a single LLM judge and need human validation. read the letter →

arxiv 2506.02708 v1 pith:KX4CIXDC submitted 2025-06-03 cs.CV cs.CL

Naoto Tanji , Toshihiko Yamasaki This is my paper
classification cs.CVcs.CL
keywords visionlanguagemodelimageaestheticassessmentscoringself-explanationDirectPreferenceOptimizationself-trainingmergingexplainableAI
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

The paper tries to establish that an instruction-tuned vision-language model can improve both its image-scoring accuracy and its natural-language justifications using only an image-score dataset and its own generated text, with no external explanation annotations or helper models. The training loop creates preference pairs by asking the model to justify a ground-truth score as the chosen response and an incorrect score as the rejected response, then applies Direct Preference Optimization. A second dataset, made by swapping the incorrect score in rejected samples back to the ground truth, pushes explanations to match the scores they accompany. The two specialized models are merged and the loop repeats. Across four vision-language models and two aesthetic-assessment datasets, iterative training raises rank correlations with human scores and raises an LLM-judge consistency score, with gains saturating around iteration four.

What carries the argument

The load-bearing object is the pair of preference datasets constructed from score-conditioned continuation. In the score dataset, chosen and rejected responses are generated by prompting the VLM with the ground-truth score versus a randomly chosen wrong score and letting it continue with an explanation, so Direct Preference Optimization makes correct-score justifications more likely. In the consistency dataset, the wrong score in each rejected sample is replaced by the ground-truth score, so chosen and rejected share the same numeric score and differ only in whether the explanation was generated to justify that score; training on this contrast rewards explanations that are genuinely consistent with the assigned score. Training the two datasets separately and merging the resulting adapters with a parameter-merging procedure that resolves sign conflicts yields one model with both capabilities, and the entire cycle is repeated with data generated by the current model.

What would settle it

Have human raters score a sample of the DPO pairs: for each image, compare the explanation generated under the ground-truth score against the explanation generated under an incorrect score, and ask which one better justifies the ground-truth score. If humans do not systematically prefer the chosen responses, or if the consistency-score gains seen with an LLM judge disappear when human ratings replace it, then the preference signal that drives the entire self-training loop is invalid.

Watch

Extended reading notes

Core claim

The central claim is that the VLM's own score-conditioned text is a sufficient training signal to bootstrap better scoring and better self-consistency. When the model is prompted with the ground-truth score and asked to justify it, the resulting explanation is treated as a chosen response; the same image with a deliberately wrong score yields a rejected response. Contrasting these pairs teaches score prediction while preserving language ability. Replacing the wrong score in rejected samples with the true score isolates the consistency signal: both responses carry the same score, so the model must learn to prefer text that actually justifies that score. Separately training one model on each dataset and merging them lets the two abilities combine, and iterating the loop with freshly generated data compounds the gain. The authors demonstrate this on AVA and AADB with four base VLMs, reporting steady SRCC improvements and consistency-score increases, with the best balanced result on AVA at iteration four reaching SRCC 0.739 and consistency 3.57, up from zero-shot values of 0.446 and 3.36.

Load-bearing premise

Everything rests on the premise that, when the model is asked to justify a ground-truth score, the resulting text is actually a better explanation of that score than text produced for an incorrect score, and the paper checks this only with an LLM judge on generated data, not with human ratings.

Editorial extensions

If this is right

  • With only image-score labels, an instruction-tuned VLM can bootstrap better scoring and better score-explanation consistency, producing all training text itself.
  • Iteration matters: static DPO data improves score accuracy but yields lower explanation consistency than the iterative loop, which regenerates the data from each updated model.
  • Score-only supervised fine-tuning destroys explanation quality, so the contrastive text-conditioned formulation is what preserves language ability while raising accuracy.
  • The merged iteration-4 model on AVA reaches SRCC 0.739 with consistency 3.57, comparable to or better than non-VLM scoring baselines, though not better than pretrained VLM scoring specialists.
  • Gains appear across model sizes, from a roughly 0.5B to a 7B vision-language model and on both datasets, with smaller improvements on AADB that the paper attributes to its smaller size.

Reading between the lines

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

  • Editorial extension: the same score-conditioning recipe should transfer to any rating-labeled perception task, such as image quality, video quality, or medical severity scoring, because it never needs domain-specific explanation text.
  • A further extension: a one-shot application of the consistency dataset is a plausible cheap alignment step for any VLM that must couple a numeric judgment with a caption, and the iterative loop may add little when the base model's text is already coherent.
  • One testable extension: the smallest model's consistency regression on AADB hints that self-generated explanations inherit the base model's zero-shot bias toward high scores, suggesting guardrails such as score-balanced data sampling or rejection sampling.
  • An interpretability caution: the reported consistency metric only checks that text plausibly justifies the assigned number, not faithfulness to the model's internal decision process, so high-stakes deployment would need an independent faithfulness audit.
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

4 major / 9 minor

Summary. The paper proposes an iterative self-training method for vision-language models (VLMs) that enables them to output an aesthetic score and a natural-language justification for that score. Using only an image-scoring dataset (AVA or AADB) and an instruction-tuned VLM, the method generates DPO training data by conditioning the model's explanation on either the ground-truth score (chosen) or an incorrect score (rejected). A second DPO dataset is constructed by replacing the incorrect score in rejected samples with the correct score, contrasting explanations that are consistent or inconsistent with the score. The model is trained on each dataset separately and the resulting LoRA adapters are merged with TIES-Merging. The process is iterated with the improved model generating the next round of data. Experiments on four VLM sizes report improved PLCC/SRCC for score prediction and improved GPT-4o-judged consistency between predicted scores and explanations relative to the zero-shot base model. The paper also includes an ablation comparing iterative DPO with SFT and non-iterative DPO, and additional analyses of dialogue ability and example outputs.

Significance. If the claims are correct, the method offers a practical way to obtain coherent self-explanations for image scoring without any external text annotations or human-written explanations, which is a valuable contribution to explainable image scoring. The score-learning half is externally grounded in human rating labels, and the DPO data construction is simple and clearly described. However, the explanation-coherence claims rest on a GPT-4o-based judge without human validation, and the score-prediction evaluation includes a post-hoc least-squares calibration on validation labels that can inflate reported correlations. Both issues affect the central 'improves both' claim and currently limit the reliability and comparability of the results.

major comments (4)
  1. [4.3] The conversion from integer score bins to original-scale predictions uses reference values \bar{s}_i that are optimized by a least-squares fit to validation data after training. Because these values are chosen to minimize error against validation ground-truth labels, the reported PLCC/SRCC are upper-biased and are not directly comparable with previous methods that use a fixed mapping (e.g., Q-ALIGN). This protocol effectively tunes part of the scoring predictor on the validation split. Please report results with a fixed mapping (such as bin means computed from the training split) and, if the calibration is retained, analyze the sensitivity of PLCC/SRCC to the reference-value choice.
  2. [3.1 and 4.3] The validity of the consistency preference pairs and the consistency evaluation both depend on GPT-4o judgments: Section 3.1 states that the alignment of the generated DPO data was 'quantitatively verified' with LLM-as-a-judge, and Section 4.3 uses GPT-4o to compute the Cons metric. No human evaluation or inter-annotator agreement is reported. Since the model is trained to increase this GPT-4o-based metric, the observed gains in Cons (e.g., Table 1, from 3.36 to 3.57) could reflect the model learning shallow textual correlates that the judge rewards rather than a genuine improvement in score-explanation coherence. The coherence half of the central claim is therefore not established. Please provide human ratings on a sample of the test set, report agreement with GPT-4o, and include error bars over multiple seeds or test subsets.
  3. [Table 2b] The claim in Section 5.1 that 'across both datasets and all models, we generally observe improved score prediction accuracy and consistency with iterative training' is contradicted by the LLaVA-interleave-0.5B row on AADB: consistency drops from 2.33 (zero-shot) to 1.36 (ite-1) and recovers only to 2.16 by ite-4, never reaching the zero-shot value. The given explanation (the zero-shot model assigns score 9 to everything and produces overly positive text) is plausible, but it demonstrates a systematic failure mode for small models and means the blanket claim should be qualified. This also raises the question of robustness when the base model's score distribution is degenerate.
  4. [Table 3] The ablation shows that the scoring-accuracy benefit of iteration is not as stated: SFT score achieves PLCC/SRCC of 0.778/0.768, and the non-iterative DPO achieves PLCC 0.742, both higher than the iterative DPO's PLCC of 0.716 (with SRCC 0.739 vs 0.733). Thus the iteration does not improve score prediction over the non-iterative baseline; it only improves consistency (3.57 vs 2.98), and that improvement depends on the unvalidated GPT-4o metric. The paper should present the trade-off between score accuracy and consistency explicitly and soften the claim that iterative training improves both abilities, at minimum distinguishing 'improvement over zero-shot' from 'improvement over non-iterative DPO'.
minor comments (9)
  1. [3.1] There is a typo in 'donotalign' which should read 'do not align'.
  2. [4.3] The downsampling of the AVA test set to 1,000 instances is only mentioned in Appendix A.1; it should be stated in the main text because it affects the reported numbers.
  3. [3.2] For the first iteration, clarify whether the consistency dataset is generated but unused, or generated only after the first score-DPO training.
  4. [Figure 3] In the consistency DPO data, the rejected sample's score is replaced with the correct score but the text was generated based on an incorrect score; the figure would be clearer if the temporal order of generation and replacement were indicated.
  5. [4.2 and A.1] The data-sampling ratios (25% of AVA training data for score training, 30% of that for consistency training) are reported only in the appendix; include them in the main text so the training setup is understandable without consulting the appendix.
  6. [Table 1] The rows labeled 'score' and 'consistency' for each iteration are the individually trained models before merging; add a note to the table caption to make that interpretation explicit.
  7. [6] The Limitations section does not acknowledge the absence of human evaluation of explanation quality; this should be added as a limitation.
  8. [1] The phrase 'without relying on external data or models' is accurate only for training-data generation; the paper uses GPT-4o for evaluation. Clarify this to avoid misunderstanding.
  9. [A.2] The judge prompt evaluates consistency, usefulness, and general quality, but only consistency is reported in the main tables; state in the prompt or main text that only the consistency subscore is used for the main results.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DPO training labels are anchored to external ground-truth scores, and no load-bearing claim reduces to its own inputs by construction.

full rationale

The paper's derivation chain is not circular by construction. The score-learning DPO dataset is built by conditioning generation on the ground-truth score versus a randomly chosen incorrect score, so the preference signal is externally grounded in human-annotated labels rather than in the model's own outputs. The consistency DPO dataset is also anchored externally: chosen responses are generated under the ground-truth score, and rejected responses are generated under an incorrect score and then relabeled with the ground-truth score; the preference is defined by the difference in conditioning, not by the evaluation metric. The claimed improvements in scoring accuracy are validated against the AVA and AADB ground-truth labels via PLCC and SRCC, which are independent of the training data construction. The explanation-coherence claim is measured with GPT-4o as an LLM judge, and the same kind of judge was used earlier to sanity-check that the generated data tend to align with scores. This is a methodological weakness -- the judge's notion of consistency is used both to motivate the data and to measure the outcome, and no human ratings are reported -- but it does not make any prediction equivalent to an input by construction, because the DPO labels do not come from the judge and the training objective does not directly optimize the judge's score. There are no load-bearing self-citations and no imported uniqueness theorems. The central result is therefore self-contained with respect to circularity, though its coherence metric would benefit from human validation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities. The free parameters are mostly evaluation calibration and data construction choices; the load-bearing assumption is that the model's own score-conditioned text provides valid preference data, which is checked only with an LLM judge.

free parameters (5)
  • Reference score values sbar_i = least squares fit on validation data
    Used to convert predicted bin probabilities to original-scale scores before computing PLCC/SRCC; post-hoc fitting can inflate accuracy.
  • Number of score bins = 10
    Score discretization into ten bins based on training-data quantiles; a hand-chosen hyperparameter.
  • Incorrect-score exclusion radius = not specified
    Rejected scores are sampled randomly while excluding values close to the correct score; the threshold is not quantified.
  • Data sampling ratios = 25% of AVA scores, 30% of consistency data; 100% for AADB
    Amounts of training data used per iteration and per objective, chosen by validation behavior.
  • Training hyperparameters = lr, DPO beta, LoRA rank per Table 4
    Selected by grid search on validation; standard but part of the method.
assumptions (4)
  • domain assumption AVA and AADB average human ratings are reliable ground-truth aesthetic scores
    All training preferences and evaluation metrics are built on these labels.
  • domain assumption GPT-4o consistency scores are a valid proxy for human-judged explanation quality
    The consistency metric is produced by an LLM judge and is used both to validate the DPO data and to measure the main result; human evaluation is deferred to future work.
  • standard math Standard DPO and TIES merging assumptions hold for LoRA-tuned VLMs
    The training and merging procedures rely on published algorithms; no formal verification is provided.
  • ad hoc to paper The base VLM's explanations conditioned on true scores are more consistent with the scores than those conditioned on wrong scores
    Central to constructing the DPO preference pairs; verified only by LLM-as-a-judge on the generated data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Iterative Self-Improvement of Vision Language Models for Image Scoring and Self-Explanation." pith.science (2026). https://pith.science/paper/KX4CIXDC

@misc{pith2026250602708,
  author       = {Pith},
  title        = {Pith review of: Iterative Self-Improvement of Vision Language Models for Image Scoring and Self-Explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KX4CIXDC}},
  note         = {Machine review of arXiv:2506.02708}
}
read the original abstract

Image scoring is a crucial task in numerous real-world applications. To trust a model's judgment, understanding its rationale is essential. This paper proposes a novel training method for Vision Language Models (VLMs) to generate not only image scores but also corresponding justifications in natural language. Leveraging only an image scoring dataset and an instruction-tuned VLM, our method enables self-training, utilizing the VLM's generated text without relying on external data or models. In addition, we introduce a simple method for creating a dataset designed to improve alignment between predicted scores and their textual justifications. By iteratively training the model with Direct Preference Optimization on two distinct datasets and merging them, we can improve both scoring accuracy and the coherence of generated explanations.

Figures

Figures reproduced from arXiv: 2506.02708 by the authors.

Figure 9
Figure 9. The result indicates that the model can provide rele [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 36 canonical work pages

  1. [1]

    Iterative Self-Improvement of Vision Language Models for Image Scoring and Self-Explanation

    INTRODUCTION Deep learning is revolutionizing image analysis, enabling au- tomated classification and scoring with enhanced accuracy and efficiency. Examples include disease detection in med- ical images, defect identification in quality control, and pre- dicting advertising effectiveness. For trustworthy applications of deep learning models in these scen...

    work page Pith review arXiv 2025

  2. [2]

    IAA datasets (e.g., A V A [3] and AADB [4]) consist of photographs scored by human annotators

    RELA TED WORK Image Aesthetic AssessmentAs an example of a high-level task where natural language explanations can be particularly valuable, we tackle image aesthetic assessment (IAA) in this work. IAA datasets (e.g., A V A [3] and AADB [4]) consist of photographs scored by human annotators. Because IAA relies on subjective human perception, generating hu...

    work page

  3. [3]

    chosen” and “rejected

    METHOD The overview of our method is illustrated in Fig. 2. We start from an instruction-tuned VLM and a target dataset that con- sists of images and scores. The first step is generating two kinds of DPO datasets; one is for learning the score predic- #Score: 8 #Explain: The gradient of warm colors in the sky creates a beautiful backdrop, enhancing the ov...

    work page

  4. [4]

    Datasets As IAA datasets, we use A V A [3] and AADB [4]

    EXPERIMENTAL SETUP 4.1. Datasets As IAA datasets, we use A V A [3] and AADB [4]. We follow the train-val-test split given by [20] for A V A, and the offi- cial one for AADB. In these datasets, each image is rated by several users with integer scores. We use average scores as ground truth labels. To treat score prediction as a language generation task, we ...

    work page 2024

  5. [5]

    Main Results Table 1 presents the results of our method applied to the LLaV A-NeXT-7B model on the A V A dataset, a part of which is also plotted in Fig

    RESULTS 5.1. Main Results Table 1 presents the results of our method applied to the LLaV A-NeXT-7B model on the A V A dataset, a part of which is also plotted in Fig. 1. Zero-shot refers to the initial model Table 1. Results of our method applied to the LLaV A-NeXT- 7B model on the A V A test dataset. For reference, we also include the accuracy of some pr...

    work page

  6. [6]

    CONCLUSIONS We have presented a training method for VLMs that allows them to generate natural language justifications for their predictions on scoring tasks. Experiments across multiple VLMs and two IAA datasets demonstrate that our approach improves IAA scoring accuracy without sacrificing language generation capabilities and enhances the consistency bet...

    work page

  7. [7]

    Opportunities and challenges in explainable artificial intelligence (xai): A survey,

    Arun Das and Paul Rad, “Opportunities and challenges in explainable artificial intelligence (xai): A survey,” arXiv preprint:2006.11371, 2020

    arXiv 2006

  8. [8]

    Di- rect preference optimization: your language model is secretly a reward model,

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn, “Di- rect preference optimization: your language model is secretly a reward model,”NeurIPS, pp. 53728–53741, 2023

    work page 2023

Show all 38 references
  1. [9]

    Ava: A large-scale database for aesthetic visual analy- sis,

    Naila Murray, Luca Marchesotti, and Florent Perronnin, “Ava: A large-scale database for aesthetic visual analy- sis,”CVPR, pp. 2408–2415, 2012

    2012

  2. [10]

    Photo aesthetics ranking network with attributes and content adaptation,

    Shu Kong, Xiaohui Shen, Zhe Lin, Radomir Mech, and Charless Fowlkes, “Photo aesthetics ranking network with attributes and content adaptation,”ECCV, 2016

    2016

  3. [11]

    Image aesthetic assessment: An experimental survey,

    Yubin Deng, Chen Change Loy, and Xiaoou Tang, “Image aesthetic assessment: An experimental survey,” IEEE Signal Processing Magazine, vol. 34, no. 4, pp. 80–106, 2017

    2017

  4. [12]

    Vila: Learning image aes- thetics from user comments with vision-language pre- training,

    Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang, “Vila: Learning image aes- thetics from user comments with vision-language pre- training,”CVPR, pp. 10041–10051, 2023

    2023

  5. [13]

    Q-align: Teach- ing LMMs for visual scoring via discrete text-defined levels,

    Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtao Zhai, and Weisi Lin, “Q-align: Teach- ing LMMs for visual scoring via discrete text-defined levels,”ICML, 2024

    2024

  6. [14]

    Chain-of-thought prompting elicits reasoning in large language models,

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al., “Chain-of-thought prompting elicits reasoning in large language models,”NeurIPS, pp. 24824–24837, 2022

    2022

  7. [15]

    Multimodal explanations: Justifying de- cisions and pointing to the evidence,

    Dong Huk Park, Lisa Anne Hendricks, Zeynep Akata, Anna Rohrbach, Bernt Schiele, Trevor Darrell, and Mar- cus Rohrbach, “Multimodal explanations: Justifying de- cisions and pointing to the evidence,”CVPR, pp. 8779– 8788, 2018

    2018

  8. [16]

    Wt5?! training text-to-text models to explain their predictions,

    Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, and Karishma Malkan, “Wt5?! training text-to-text models to explain their predictions,” arXiv preprint arXiv:2004.14546, 2020

    2004 arXiv

  9. [17]

    Measuring association between labels and free-text ra- tionales,

    Sarah Wiegreffe, Ana Marasovi ´c, and Noah A Smith, “Measuring association between labels and free-text ra- tionales,”EMNLP, pp. 10266–10284, 2021

    2021

  10. [18]

    Few-shot self-rationalization with natural language prompts,

    Ana Marasovi ´c, Iz Beltagy, Doug Downey, and Matthew E Peters, “Few-shot self-rationalization with natural language prompts,”Findings of the Association for Computational Linguistics: NAACL, pp. 410–424, 2022

    2022

  11. [19]

    Reframing human- ai collaboration for generating free-text explanations,

    Sarah Wiegreffe, Jack Hessel, Swabha Swayamdipta, Mark Riedl, and Yejin Choi, “Reframing human- ai collaboration for generating free-text explanations,” NAACL-HLT, pp. 632–658, 2022

    2022

  12. [20]

    Large language models can self-improve,

    Jiaxin Huang, Shixiang Gu, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu, and Jiawei Han, “Large language models can self-improve,”EMNLP, pp. 1051–1068, 2023

    2023

  13. [21]

    Self-refine: iterative refinement with self-feedback,

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al., “Self-refine: iterative refinement with self-feedback,” NeurIPS, pp. 46534–46594, 2023

    2023

  14. [22]

    Self-rewarding language models,

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason E Weston, “Self-rewarding language models,”ICML, 2024

    2024

  15. [23]

    Iterative reasoning preference optimization,

    Richard Yuanzhe Pang, Weizhe Yuan, He He, Kyunghyun Cho, Sainbayar Sukhbaatar, and Jason We- ston, “Iterative reasoning preference optimization,” NeurIPS, pp. 116617–116637, 2024

    2024

  16. [24]

    Enhancing large vision language models with self-training on image comprehension,

    Yihe Deng, Pan Lu, Fan Yin, Ziniu Hu, Sheng Shen, Quanquan Gu, James Zou, Kai-Wei Chang, and Wei Wang, “Enhancing large vision language models with self-training on image comprehension,”NeurIPS, pp. 131369–131397, 2024

    2024

  17. [25]

    Ties-merging: resolving in- terference when merging models,

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal, “Ties-merging: resolving in- terference when merging models,”NeurIPS, pp. 7093– 7115, 2023

    2023

  18. [26]

    Nima: Neural image assessment,

    Hossein Talebi and Peyman Milanfar, “Nima: Neural image assessment,”IEEE Transactions on Image Pro- cessing, vol. 27, no. 8, pp. 3998–4011, 2018

    2018

  19. [27]

    Llava- next-interleave: Tackling multi-image, video, and 3d in large multimodal models,

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li, “Llava- next-interleave: Tackling multi-image, video, and 3d in large multimodal models,”CoRR, 2024

    2024

  20. [28]

    Internvl: Scaling up vi- sion foundation models and aligning for generic visual- linguistic tasks,

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al., “Internvl: Scaling up vi- sion foundation models and aligning for generic visual- linguistic tasks,”CVPR, pp. 24185–24198, 2024

    2024

  21. [29]

    Improved baselines with visual instruction tun- ing,

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee, “Improved baselines with visual instruction tun- ing,”CVPR, pp. 26296–26306, 2024

    2024

  22. [30]

    Llava- next: Improved reasoning, ocr, and world knowl- edge,

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuan- han Zhang, Sheng Shen, and Yong Jae Lee, “Llava- next: Improved reasoning, ocr, and world knowl- edge,”https://llava-vl.github.io/blog/ 2024-01-30-llava-next, 2024

    2024

  23. [31]

    LoRA: Low-rank adaptation of large language models,

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen, “LoRA: Low-rank adaptation of large language models,”ICLR, 2022

    2022

  24. [32]

    Musiq: Multi-scale image quality transformer,

    Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang, “Musiq: Multi-scale image quality transformer,”ICCV, pp. 5148–5157, 2021. Table 4. Hyperparameters used in our training. dataset model lr DPOβ LoRA rank A V A LLaV A-NeXT-7B 5.e-5 0.1 64 LLaV A-1.5-7B 3.e-5 0.2...

    2021

  25. [33]

    This could be achieved by using a higher megapixel camera or a camera with better low-light capabilities

    **Resolution and Clarity**: Increase the resolution of the image to capture more detail and make the soup and its textures more visible. This could be achieved by using a higher megapixel camera or a camera with better low-light capabilities

  26. [34]

    Natural light is often preferred for food photography, as it can create a more appealing and natural look

    **Lighting**: Ensure that the lighting is even and well- balanced to avoid harsh shadows and highlights. Natural light is often preferred for food photography, as it can create a more appealing and natural look

  27. [35]

    Consider the rule of thirds and the use of negative space to create a more balanced and visually appealing image

    **Composition**: Pay attention to the composition of the shot. Consider the rule of thirds and the use of negative space to create a more balanced and visually appealing image

  28. [36]

    This can help draw attention to the main subject of the image

    **Depth of Field**: Use a shallower depth of field to keep the soup in focus while blurring the background slightly. This can help draw attention to the main subject of the image

  29. [37]

    This can be done using photo editing software

    **Color and Contrast**: Adjust the color balance and contrast to enhance the colors and textures in the soup. This can be done using photo editing software

  30. [38]

    This could include adjusting the exposure, contrast, and saturation to make the colors pop and the image more vibrant

    **Post-Processing**: Use photo editing software to enhance the image further. This could include adjusting the exposure, contrast, and saturation to make the colors pop and the image more vibrant. By implementing these suggestions, the image could be improved, resulting in a h...

Pith tools

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