In this workshop, we will discuss how knowledge about the visual world is represented in modern generative models for images, videos, and 3D assets. Recent advances in generative modeling have been successful in creating rich, diverse and increasingly convincing photorealistic and stylized images. The workshop aims to investigate how these models internally represent and process visual information, and whether they bring us closer to fulfillment of the known mantra “vision is inverse graphics”. We seek to understand how well generative models, such as GANs, auto-regressive models, and diffusion models, comprehend semantic constructs that are commonly used to convey visual understanding, such as object recognition, scene understanding, spatial awareness, intrinsic image decomposition, and so on. Can this understanding be leveraged to solve inverse (recognition) problems? Can it be improved to further enhance generative models’ abilities? Is something important still missing from how our large models represent the visual world?

• Internal Representation of Visual Knowledge: Analyzing how generative models internally represent visual information, including object features, textures, and spatial relationships, and 3D geometry

• Implicit Computer Vision Tasks in Generative Models: Investigating whether, and how generative models perform implicit computer vision tasks during image synthesis, such as object placement, lighting consistency, and perspective rendering.

• Knowledge Gaps in Generative Models: What Visual Knowledge today's Generative Models lack but should be aware of? How to bridge this knowledge gap? 

• Transfer of Knowledge from Generative to Discriminative Models and vice-versa: Exploring methods to transfer learned representations between generative to discriminative models for tasks like object recognition and scene segmentation. 

• Evaluating Computer Vision Knowledge in Generative Models: Developing metrics and methodologies to evaluate the extent and fidelity of computer vision knowledge within generative models.

We have stopped accepting new submissions. We'll contact authors who have already filled the interest form very soon.

We are collecting interest for poster presentations at our workshop. If you are interested in presenting your work, please fill out this Google form. We will follow up on the status of the poster presentations in the coming months.

📧  Contact: Anand Bhattad (bhattad@ttic.edu)