Visualizing poetry with deep semantic understanding and consistency evaluation

Poetry-to-image generation aims to translate abstract poetic language into visual art, but existing methods fail to capture deep emotions and artistic essence, producing images that lack the poem’s aesthetic quality. To address this, we propose a multi-stage framework that deeply understands poetic semantics. Our method first extracts key emotional, imagery, and rhetorical features from poetry. These multi-dimensional features are then fused to guide the generation of detailed image prompts. Finally, a poem-image consistency module evaluates the output, ensuring the final image is highly aligned with the poem’s content and style. Experiments on our new “Poetic Visions” dataset show our method significantly outperforms existing approaches across multiple metrics, including IS, FID, CLIP Score, and human evaluation, producing images with superior artistic expression and semantic consistency.