In text-to-image in-context learning (T2I-ICL), a model must infer a latent compositional pattern from few-shot demonstrations to generate a query image. Recent studies show that state-of-the-art multimodal large language models struggle in this setting, primarily due to limited compositional reasoning and sensitivity to prompt construction. To address this, we propose a Tree-of-Thoughts (ToT) reasoning framework that introduces a multi-stage reasoning and selection layer, generating, evaluating, and selecting among multiple candidate hypotheses before constructing the final prompt for image synthesis.
By exploring alternative reasoning branches and selecting a coherent interpretation, our approach mitigates prompt ambiguity and compositional errors. We implement this approach in a complete ToT-T2I-ICL inference pipeline and evaluate it on the CoBSAT benchmark. Both qualitative and quantitative results show that structured multi-branch reasoning leads to more consistent and semantically aligned image generation compared to baseline and Chain-of-Thought prompting strategies, without any additional training or fine-tuning.
Blogger's Review: This paper significantly enhances text-to-image generation through the introduction of the Tree-of-Thoughts reasoning framework, showcasing the potential of multi-stage reasoning in addressing compositional reasoning challenges, making it a valuable contribution to the field of multimodal learning.