Text-to-image diffusion models have demonstrated significant capabilities to generate diverse and detailed visuals in various domains, and story visualization is emerging as a particularly promising application. However, as their use in real-world creative domains increases, the need for providing enhanced control, refinement, and the ability to modify images post-generation in a consistent manner becomes an important challenge. Existing methods often lack the flexibility to apply fine or coarse edits while maintaining visual and narrative consistency across multiple frames, preventing creators from seamlessly crafting and refining their visual stories. To address these challenges, we introduce Plot'n Polish, a zero-shot framework that enables consistent story generation and provides fine-grained control over story visualizations at various levels of detail.

Savya Khosla, Aditi Tiwari, Kushal Kafle, Simon Jenni, Handong Zhao, John Collomosse, Jing Shi. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Traditional alignment methods for Large Vision and Language Models (LVLMs) primarily rely on human-curated preference data. Human-generated preference data is costly; machine-generated preference data is limited in quality; and self-supervised preference data often introduces hallucinations. To overcome these limitations, we propose a novel Panel-of-Peers learning framework inspired by collaborative learning among humans. This approach leverages a panel of LVLMs, each evaluating and learning from their collective outputs through an iterative self-improvement process. By simulating a peer review system, our models generate, assess, and refine outputs in response to a curated set of prompts, mimicking a classroom learning environment. We demonstrate that this methodology enhances model performance without requiring extensive human-labeled datasets. Our experiments show significant improvement across multiple benchmarks, demonstrating the potential of peer evaluations as a scalable alternative to self-supervised alignment. Notably, we show that Panel-of-Peers increases the average score on fifteen benchmarks from 48% to 57%

Multimodal large language models (MLLMs) are increasingly used to evaluate text-to-image (TTI) generation systems, providing automated judgments based on visual and textual context. However, these "judge" models often suffer from biases, overconfidence, and inconsistent performance across diverse image domains. While prompt ensembling has shown promise for mitigating these issues in unimodal, text-only settings, our experiments reveal that standard ensembling methods fail to generalize effectively for TTI tasks. To address these limitations, we propose a new multimodal-aware method called Multimodal Mixture-of-Bayesian Prompt Ensembles (MMB). Our method uses a Bayesian prompt ensemble approach augmented by image clustering, allowing the judge to dynamically assign prompt weights based on the visual characteristics of each sample. We show that MMB improves accuracy in pairwise preference judgments and greatly enhances calibration, making it easier to gauge the judge's true uncertainty. In evaluations on two TTI benchmarks, HPSv2 and MJBench, MMB outperforms existing baselines in alignment with human annotations and calibration across varied image content. Our findings highlight the importance of multimodal-specific strategies for judge calibration and suggest a promising path forward for reliable large-scale TTI evaluation.

Recent advancements in Vision-Language Models (VLMs) have marked a significant leap in bridging the gap between computer vision and natural language processing. However, traditional VLMs, trained through contrastive learning on limited and noisy image-text pairs, often lack the spatial and linguistic understanding to generalize well to dense vision tasks or less common languages. Our approach, Solid Foun-dation CLIP (SF-CLIP), circumvents this issue by implicitly building on the solid visual and language understanding of foundational models trained on vast amounts of unimodal data. SF-CLIP integrates contrastive image-text pretraining with a masked knowledge distillation from large foundational text and vision models. This methodology guides our VLM in developing robust text and image representations. As a result, SF-CLIP shows exceptional zero-shot classification accuracy and enhanced image and text retrieval capabilities, setting a new state of the art for ViT-B/16 trained on YFCC15M and CC12M.Moreover, the dense per-patch supervision enhances our zero-shot and linear probe performance in semantic segmentation tasks. A remarkable aspect of our model is its multilingual proficiency, evidenced by strong retrieval results in multiple languages despite being trained predominantly on English data. We achieve all of these improvements without sacrificing the training efficiency through our selective application of masked distillation and the inheritance of teacher word embeddings.

Recent dataset deduplication techniques have demonstrated that content-aware dataset pruning can dramatically reduce the cost of training Vision-Language Pre-trained (VLP) models without significant performance losses compared to training on the original dataset. These results have been based on pruning commonly used image-caption datasets collected from the web - datasets that are known to harbor harmful social biases that may then be codified in trained models. In this work, we evaluate how deduplication affects the prevalence of these biases in the resulting trained models and introduce an easy-to-implement modification to the recent SemDeDup algorithm that can reduce the negative effects that we observe. When examining CLIP-style models trained on deduplicated variants of LAION-400M, we find our proposed FairDeDup algorithm consistently leads to improved fairness metrics over SemDeDup on the FairFace and FACET datasets while maintaining zero-shot performance on CLIP benchmarks.

In recent years, the use of CLIP (Contrastive Language-Image Pre-Training) has become increasingly popular in a wide range of downstream applications, including zero-shot image classification and text-to-image synthesis. Despite being trained on a vast dataset, the CLIP model has been found to exhibit biases against certain protected attributes, such as gender and race. While previous research has focused on the impact of such biases on image classification, there has been little investigation into their effects on CLIP-based generative tasks. In this paper, we aim to address this gap in the literature by uncovering the queries for which the CLIP model introduces biases in the text-based image editing task. Through a series of experiments, we demonstrate that these biases can have a significant impact on the quality and content of the generated images. To mitigate these biases, we propose a debiasing technique that does not require retraining either the CLIP model or the underlying generative model. Our results show that our proposed framework can effectively reduce the impact of biases in CLIP-based image editing models. Overall, this paper highlights the importance of addressing biases in CLIP-based generative tasks and provides practical solutions that can be readily adopted by researchers and practitioners working in this area.1

We propose Subject-Conditional Relation Detection (SCoRD), where conditioned on an input subject, the goal is to predict all its relations to other objects in a scene along with their locations. Based on the Open Images dataset, we propose a challenging OIv6-SCoRD benchmark such that the training and testing splits have a distribution shift in terms of the occurrence statistics of subject, relation, object triplets. To solve this problem, we propose an auto-regressive model that given a subject, it predicts its relations, objects, and object locations by casting this output as a sequence of tokens. First, we show that previous scene-graph prediction methods fail to produce as exhaustive an enumeration of relation-object pairs when conditioned on a subject on this benchmark. Particularly, we obtain a recall@3 of 83.8% for our relation-object predictions compared to the 49.75% obtained by a recent scene graph detector. Then, we show improved generalization on both relation-object and object-box predictions by leveraging during training relation-object pairs obtained automatically from textual captions and for which no object-box annotations are available. Particularly, for subject, relation, object triplets for which no object locations are available during training, we are able to obtain a recall@3 of 33.80% for relation-object pairs and 26.75% for their box locations.

We propose a margin-based loss for tuning joint vision-language models so that their gradient-based explanations are consistent with region-level annotations provided by humans for relatively smaller grounding datasets. We refer to this objective as Attention Mask Consistency (AMC) and demonstrate that it produces superior visual grounding results than previous methods that rely on using vision-language models to score the outputs of object detectors. Particularly, a model trained with AMC on top of standard vision-language modeling objectives obtains a state-of-the-art accuracy of 86.49% in the Flickr30k visual grounding benchmark, an absolute improvement of 5.38% when compared to the best previous model trained under the same level of supervision. Our approach also performs exceedingly well on established benchmarks for referring expression comprehension where it obtains 80.34% accuracy in the easy test of RefCOCO+, and 64.55% in the difficult split. AMC is effective, easy to implement, and is general as it can be adopted by any vision-language model, and can use any type of region annotations.

A critical problem in deep learning is that systems learn inappropriate biases, resulting in their inability to perform well on minority groups. This has led to the creation of multiple algorithms that endeavor to mitigate bias. However, it is not clear how effective these methods are. This is because study protocols differ among papers, systems are tested on datasets that fail to test many forms of bias, and systems have access to hidden knowledge or are tuned specifically to the test set. To address this, we introduce an improved evaluation protocol, sensible metrics, and a new dataset, which enables us to ask and answer critical questions about bias mitigation algorithms. We evaluate seven state-of-the-art algorithms using the same network architecture and hyperparameter selection policy across three benchmark datasets. We introduce a new dataset called Biased MNIST that enables assessment of robustness to multiple bias sources. We use Biased MNIST and a visual question answering (VQA) benchmark to assess robustness to hidden biases. Rather than only tuning to the test set distribution, we study robustness across different tuning distributions, which is critical because for many applications the test distribution may not be known during development. We find that algorithms exploit hidden biases, are unable to scale to multiple forms of bias, and are highly sensitive to the choice of tuning set. Based on our findings, we implore the community to adopt more rigorous assessment of future bias mitigation methods. All data, code, and results are publicly available1.

Visual attributes constitute a large portion of information contained in a scene. Objects can be described using a wide variety of attributes which portray their visual appearance (color, texture), geometry (shape, size, posture), and other intrinsic properties (state, action). Existing work is mostly limited to study of attribute prediction in specific domains. In this paper, we introduce a large-scale in-the-wild visual attribute prediction dataset consisting of over 927K attribute annotations for over 260K object instances. Formally, object attribute prediction is a multi-label classification problem where all attributes that apply to an object must be predicted. Our dataset poses significant challenges to existing methods due to large number of attributes, label sparsity, data imbalance, and object occlusion. To this end, we propose several techniques that systematically tackle these challenges, including a base model that utilizes both low- and high-level CNN features with multi-hop attention, reweighting and resampling techniques, a novel negative label expansion scheme, and a novel supervised attribute-aware contrastive learning algorithm. Using these techniques, we achieve near 3.7 mAP and 5.7 overall F1 points improvement over the current state of the art. Further details about the VAW dataset can be found at https://vawdataset.com/

Existing Visual Question Answering (VQA) methods tend to exploit dataset biases and spurious statistical correlations, instead of producing right answers for the right reasons. To address this issue, recent bias mitigation methods for VQA propose to incorporate visual cues (e.g., human attention maps) to better ground the VQA models, showcasing impressive gains. However, we show that the performance improvements are not a result of improved visual grounding, but a regularization effect which prevents over-fitting to linguistic priors. For instance, we find that it is not actually necessary to provide proper, human-based cues; random, insensible cues also result in similar improvements. Based on this observation, we propose a simpler regularization scheme that does not require any external annotations and yet achieves near state-of-the-art performance on VQA-CPv2.

Out-of-distribution (OOD) testing is increasingly popular for evaluating a machine learning system's ability to generalize beyond the biases of a training set. OOD benchmarks are designed to present a different joint distribution of data and labels between training and test time. VQA-CP has become the standard OOD benchmark for visual question answering, but we discovered three troubling practices in its current use. First, most published methods rely on explicit knowledge of the construction of the OOD splits. They often rely on inverting'' the distribution of labels, e.g. answering mostlyyes'' when the common training answer was ``no''. Second, the OOD test set is used for model selection. Third, a model's in-domain performance is assessed after retraining it on in-domain splits (VQA v2) that exhibit a more balanced distribution of labels. These three practices defeat the objective of evaluating generalization, and put into question the value of methods specifically designed for this dataset. We show that embarrassingly-simple methods, including one that generates answers at random, surpass the state of the art on some question types. We provide short- and long-term solutions to avoid these pitfalls and realize the benefits of OOD evaluation.

Chart question answering (CQA) is a newly proposed visual question answering (VQA) task where an algorithm must answer questions about data visualizations, e.g. bar charts, pie charts, and line graphs. CQA requires capabilities that natural-image VQA algorithms lack: fine-grained measurements, optical character recognition, and handling out-of-vocabulary words in both questions and answers. Without modifications, state-of-the-art VQA algorithms perform poorly on this task. Here, we propose a novel CQA algorithm called parallel recurrent fusion of image and language (PReFIL). PReFIL first learns bimodal embeddings by fusing question and image features and then intelligently aggregates these learned embeddings to answer the given question. Despite its simplicity, PReFIL greatly surpasses state-of-the art systems and human baselines on both the FigureQA and DVQA datasets. Additionally, we demonstrate that PReFIL can be used to reconstruct tables by asking a series of questions about a chart.

Most counting questions in visual question answering (VQA) datasets are simple and require no more than object detection. Here, we study algorithms for complex counting questions that involve relationships between objects, attribute identification, reasoning, and more. To do this, we created TallyQA, the world’s largest dataset for open-ended counting. We propose a new algorithm for counting that uses relation networks with region proposals. Our method lets relation networks be efficiently used with high-resolution imagery. It yields stateof-the-art results compared to baseline and recent systems on both TallyQA and the HowMany-QA benchmark.

Visual Question Answering (VQA) research is split into two camps: the first focuses on VQA datasets that require natural image understanding and the second focuses on synthetic datasets that test reasoning. A good VQA algorithm should be capable of both, but only a few VQA algorithms are tested in this manner. We compare five state-of-the-art VQA algorithms across eight VQA datasets covering both domains. To make the comparison fair, all of the models are standardized as much as possible, e.g., they use the same visual features, answer vocabularies, etc. We find that methods do not generalize across the two domains. To address this problem, we propose a new VQA algorithm that rivals or exceeds the state-of-the-art for both domains.

Bar charts are an effective way to convey numeric information, but today's algorithms cannot parse them. Existing methods fail when faced with even minor variations in appearance. Here, we present DVQA, a dataset that tests many aspects of bar chart understanding in a question answering framework. Unlike visual question answering (VQA), DVQA requires processing words and answers that are unique to a particular bar chart. State-of-the-art VQA algorithms perform poorly on DVQA, and we propose two strong baselines that perform considerably better. Our work will enable algorithms to automatically extract numeric and semantic information from vast quantities of bar charts found in scientific publications, Internet articles, business reports, and many other areas.

In visual question answering (VQA), an algorithm must answer text-based questions about images. While multiple datasets for VQA have been created since late 2014, they all have flaws in both their content and the way algorithms are evaluated on them. As a result, evaluation scores are inflated and predominantly determined by answering easier questions, making it difficult to compare different methods. In this paper, we analyze existing VQA algorithms using a new dataset called the Task Driven Image Understanding Challenge (TDIUC), which has over 1.6 million questions organized into 12 different categories. We also introduce questions that are meaningless for a given image to force a VQA system to reason about image content. We propose new evaluation schemes that compensate for over-represented question-types and make it easier to study the strengths and weaknesses of algorithms. We analyze the performance of both baseline and state-of-the-art VQA models, including multi-modal compact bilinear pooling (MCB), neural module networks, and recurrent answering units. Our experiments establish how attention helps certain categories more than others, determine which models work better than others, and explain how simple models (e.g. MLP) can surpass more complex models (MCB) by simply learning to answer large, easy question categories.

Data augmentation is widely used to train deep neural networks for image classification tasks.Simply flipping images can help learning by increasing the number of training images by a factor of two.However, data augmentation in natural language processing is much less studied.Here, we describe two methods for data augmentation for Visual Question Answering (VQA).The first uses existing semantic annotations to generate new questions.The second method is a generative approach using recurrent neural networks.Experiments show the proposed schemes improve performance of baseline and state-of-the-art VQA algorithms.

Recently, algorithms for object recognition and related tasks have become sufficiently proficient that new vision tasks can now be pursued. In this paper, we build a system capable of answering open-ended text-based questions about images, which is known as Visual Question Answering (VQA). Our approach's key insight is that we can predict the form of the answer from the question. We formulate our solution in a Bayesian framework. When our approach is combined with a discriminative model, the combined model achieves state-of-the-art results on four benchmark datasets for open-ended VQA: DAQUAR, COCO-QA, The VQA Dataset, and Visual7W.