Thanks AK for sharing our work! By utilizing discriminative models of various visual tasks as reward models, we significantly improve the controllability of image generation under various conditions.

🔬 Introducing SciVideoBench: a new benchmark for scientific video reasoning in large multimodal models 🎥 1,000 research-level questions across 241 experimental videos spanning physics, biology, chemistry & medicine 🧠 It challenges models on conceptual, hypothetical, and