Search papers, labs, and topics across Lattice.
This paper investigates the impact of diffusion-based inpainting artifacts on the language generation capabilities of vision-language models. They employ a two-stage diagnostic setup, reconstructing masked image regions and then using the inpainted images for captioning, comparing the resulting captions to those generated from original images. The study finds a consistent correlation between pixel-level and perceptual reconstruction metrics and the lexical and semantic quality of generated captions, indicating that inpainting artifacts negatively impact downstream language tasks.
Inpainting artifacts in images don't just degrade visual quality; they systematically warp the language generated by vision-language models.
We study how visual artifacts introduced by diffusion-based inpainting affect language generation in vision-language models. We use a two-stage diagnostic setup in which masked image regions are reconstructed and then provided to captioning models, enabling controlled comparisons between captions generated from original and reconstructed inputs. Across multiple datasets, we analyze the relationship between reconstruction fidelity and downstream caption quality. We observe consistent associations between pixel-level and perceptual reconstruction metrics and both lexical and semantic captioning performance. Additional analysis of intermediate visual representations and attention patterns shows that inpainting artifacts lead to systematic, layer-dependent changes in model behavior. Together, these results provide a practical diagnostic framework for examining how visual reconstruction quality influences language generation in multimodal systems.
