CoF-T2I: Video Models as Pure Visual Reasoners for Text-to-Image Generation

Recent video generation models have revealed the emergence of Chain-of-Frame (CoF) reasoning, enabling frame-by-frame visual inference. With this capability, video models have been successfully applied to various visual tasks (e.g., maze solving, visual puzzles). However, their potential to enhance text-to-image (T2I) generation remains largely unexplored due to the absence of a clearly defined visual reasoning starting point and interpretable intermediate states in the T2I generation process. To bridge this gap, we propose CoF-T2I, a model that integrates CoF reasoning into T2I generation via progressive visual refinement, where intermediate frames act as explicit reasoning steps and the final frame is taken as output. To establish such explicit generation process, we curate CoF-Evol-Instruct, a dataset of CoF trajectories that model the generation process from semantics to aesthetics. To further improve quality and avoid motion artifacts, we enable an independent encoding operation for each frame. Experiments show that CoF-T2I significantly outperforms the base video model and achieves competitive performance, reaching 0.86 on GenEval and 7.468 on Imagine-Bench. These results indicate the substantial promise of video models for advancing high-quality text-to-image generation.

Text-to-image AI systems can create realistic pictures from written descriptions, but they often make mistakes when the description contains many details, such as multiple objects or specific relationships between them. People usually handle such tasks by imagining a rough version first and then improving it step by step. Inspired by this idea, we propose CoF-T2I, a method that uses a video generation model to create a short visual refinement process for image generation. The model first produces a rough image, then a more accurate version, and finally a polished image. We keep only the final image, while using the earlier frames to guide the generation process. This helps the model follow complex prompts more accurately without needing an extra verifier or written reasoning steps. Our work suggests that video models can be useful not only for making videos, but also for improving how AI systems generate still images.