Approaches Used by State-of-the-Art Vision-Language Models for Handling High-Resolution Images

Author(s): Duci Nguyen

Originally published on Towards AI.

This member-only story is on us. Upgrade to access all of Medium.

Have you ever failed to ask a Vision-Language Model (VLM) to search for specific objects or thoroughly explain details about a high-resolution image?

Image by V* Visual Search paper

When I studied about multimodal large language models (MLLMs), I face a significant challenge: these models depend heavily on pre-trained vision encoders like the CLIP image encoder, which are trained on smaller-scale images — typically 224×224 or 336×336 pixels. This training approach poses a problem when deploying these models, as images are often downscaled to similar resolutions, resulting in the loss of crucial details in higher-resolution images.

Therefore in this article, I will introduce the four most recent methods and provide a middle-level explanation of each method to ensure clarity and accessibility for all readers.

Image by FlexAttention paper

(a) Using Low-Resolution Vision-Language Models (VLMs): I find that these models downsample high-resolution images to meet vision encoder standards. This often results in significant detail loss, which hampers the model’s ability to accurately address queries about the image.

(b) Employing High-Resolution VLMs: While these models handle high-resolution images directly, they consume a large number of image tokens, leading to excessive computational demands and inefficiency.

4.1 FlexAttention for Efficient… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI