Suhail Doshi: The Future of Computer Vision

Unified Vision

Suhail Doshi is pioneering the development of a unified computer vision model that aims to revolutionize how we interact with digital media. He highlights the underinvestment in the vision category compared to language models, emphasizing the potential for a graphics model capable of thousands of tasks, similar to GPT-4 in language processing 1. Suhail's vision includes creating a model that can generate, edit, and understand pixels, pushing the boundaries of current graphics capabilities 2.

The end goal is to create a graphics model that's capable of, like, thousands of tasks.

--- Suhail Doshi

He believes that such a model would significantly enhance the aesthetic quality of images and address existing bottlenecks in current architectures 3.

   

Image Quality

Enhancing image quality through improved contrast and color vibrancy is a key focus for Suhail Doshi. He notes that users appreciate vibrant colors, yet many models struggle with simple tasks like producing a black background 4. Suhail's approach involves innovative techniques like offset noise and EDM to achieve rich color and contrast, aiming to surpass existing models like SDXL 5.

Users really love the vibrancy of colors.

--- Suhail Doshi

He emphasizes the importance of user feedback in refining these models, ensuring they meet diverse aesthetic preferences 6.

   

Benchmarks

Benchmarks play a crucial role in the development of AI models, yet Suhail Doshi argues they often fail to capture the full potential of these technologies. He believes that companies that create specific benchmarks aligned with user needs will produce models with higher utility 7. Suhail critiques the current state of evaluations, noting that many benchmarks are not prepared for the sophistication of emerging models 8.

Benchmarks are not prepared for how powerful these models will become.

--- Suhail Doshi

He advocates for a shift towards more qualitative assessments that better reflect real-world applications and user experiences 9.