We present the first study on how Multimodal LLMs' (MLLMs) reasoning ability shall be elicited to evaluate the aesthetics of artworks. To facilitate this investigation, we construct MM-StyleBench, a novel high-quality dataset for benchmarking artistic stylization. We then develop a principled method for human preference modeling and perform a systematic correlation analysis between MLLMs' responses and human preference. Our experiments reveal an inherent hallucination issue of MLLMs in art evaluation, associated with response subjectivity. ArtCoT is proposed, demonstrating that art-specific task decomposition and the use of concrete language boost MLLMs' reasoning ability for aesthetics. Our findings offer valuable insights into MLLMs for art and can benefit a wide range of downstream applications, such as style transfer and artistic image generation. Code available at https://github.com/songrise/MLLM4Art.

Prompt-tuning has demonstrated parameter-efficiency in fusing unimodal foundation models for multimodal tasks. However, its limited adaptivity and expressiveness lead to suboptimal performance when compared with other tuning methods. In this paper, we address this issue by disentangling the vanilla prompts to adaptively capture dataset-level and instance-level features. Building upon this disentanglement, we introduce the mixture of prompt experts (MoPE) technique to enhance expressiveness. MoPE leverages multimodal pairing priors to route the most effective prompt on a per-instance basis. Compared to vanilla prompting, our MoPE-based conditional prompting exhibits greater expressiveness for multimodal fusion, scaling better with the training data and the overall number of trainable parameters. We also study a regularization term for expert routing, leading to emergent expert specialization, where different experts focus on different concepts, enabling interpretable soft prompting. Extensive experiments across three multimodal datasets demonstrate that our method achieves state-of-the-art results, matching or even surpassing the performance of fine-tuning, while requiring only 0.8% of the trainable parameters. Code will be released: https://github.com/songrise/MoPE.

We show that the representation of one modality can effectively guide the prompting of another modality for parameter-efficient multimodal fusion. Specifically, we first encode one modality and use its representation as a prior to conditionally prompt all frozen layers of the other modality. This is achieved by disentangling the vanilla prompt vectors into three types of specialized prompts that adaptively capture global-level and instance-level features. To better produce the instance-wise prompt, we introduce the mixture of prompt experts (MoPE) to dynamically route each instance to the most suitable prompt experts for encoding. We further study a regularization term to avoid degenerated prompt expert routing. Thanks to our design, our method can effectively transfer the pretrained knowledge in unimodal encoders for downstream multimodal tasks. Compared with vanilla prompting, we show that our MoPE-based conditional prompting is more expressive, thereby scales better with training data and the total number of prompts. We also demonstrate that our prompt tuning is architecture-agnostic, thereby offering high modularity. Extensive experiments over three multimodal datasets demonstrate state-of-the-art results, matching or surpassing the performance achieved through fine-tuning, while only necessitating 0.7% of the trainable parameters. Code will be released: https://github.com/songrise/ConditionalPrompt.

Recent advances in visual-language models have shown remarkable zero-shot text-image matching ability that is transferable to downstream tasks such as object detection and segmentation. Adapting these models for object counting, however, remains a formidable challenge. In this study, we first investigate transferring vision-language models (VLMs) for class-agnostic object counting. Specifically, we propose CLIP-Count, the first end-to-end pipeline that estimates density maps for open-vocabulary objects with text guidance in a zero-shot manner. To align the text embedding with dense visual features, we introduce a patch-text contrastive loss that guides the model to learn informative patch-level visual representations for dense prediction. Moreover, we design a hierarchical patch-text interaction module to propagate semantic information across different resolution levels of visual features. Benefiting from the full exploitation of the rich image-text alignment knowledge of pretrained VLMs, our method effectively generates high-quality density maps for objects-of-interest. Extensive experiments on FSC-147, CARPK, and ShanghaiTech crowd counting datasets demonstrate state-of-the-art accuracy and generalizability of the proposed method. Code is available: https://github.com/songrise/CLIP-Count.