🔥 Core Discovery: Visual generation capability naturally arises from understanding! With just 200K samples + co-training, LLMs can be taught to generate visual embeddings without extensive ...