Latent reasoning boosts image generation quality and scalability
Latent Reasoning in Continuous Space for Unified Multimodal Models
Computer Vision and Pattern RecognitionComputation and LanguageMachine Learning
Summary
Drawing pictures from text is hard because language thinking happens in words, but images are made from continuous data. The authors propose a way to think in a continuous space instead of words, which better matches how images are created. They train models in two steps: first teaching them gradually to think this way, then letting them improve by themselves. This new method improves how well pictures match descriptions and how detailed and controlled the results are, especially when using large models.
What this means in practice
- •For ai product developers: Build text-to-image generation tools that produce higher-quality and more accurate images by using continuous latent reasoning instead of token-based methods.$Commercial implications: Enables creation of advanced image generation products with better control and scaling, useful for commercial creative applications.
- •For machine learning engineers: Improve training of multimodal AI models by applying gradual curriculum fine-tuning and reinforcement learning in continuous latent spaces to enhance image generation fidelity.
Authors
Byungwoo Jeon, Yoonwoo Jeong, Hyunseok Lee, Minsu Cho, Jinwoo Shin
Abstract
Unified Multimodal Models (UMMs) achieve remarkable success on diverse multimodal understanding and generation tasks, yet they still struggle to effectively leverage language-based chain-of-thought reasoning for image generation. As a result, existing reasoning-augmented UMMs often fail to inherit the test-time scaling and fine-grained controllability observed in large language models. We attribute this failure to a fundamental modality mismatch: reasoning is performed in a discrete token space, whereas image generation operates in a continuous space. To bridge this gap, we introduce a Latent Reasoning framework in Continuous space (LARC), which alleviates the token-level bottleneck by reasoning over continuous hidden states. Specifically, the training framework consists of two stages: (i) curriculum supervised fine-tuning (SFT), and (ii) information-gain reinforcement learning (RL). First, the curriculum SFT stage gradually converts token-level reasoning into latent reasoning. Subsequently, the information-gain RL stage self-evolves latent reasoning without ground truth supervision. Experiments across diverse text-to-image generation benchmarks show that LARC consistently improves generation quality, prompt following, and compositional fidelity. Notably, LARC achieves state-of-the-art performance on GenEval with a 7%p gain over BAGEL and demonstrates clear test-time scaling behavior, outperforming token-based generation under both parallel and sequential test-time scaling strategies.