Abstract: Highlights•Zero-shot prediction is a hallmark of image reconstruction under limited data.•Text-guided methods fail to reconstruct novel images in a well-designed dataset.•Limited dataset diversity causes model outputs to collapse onto known images.•Text features do not recover an image, promoting hallucinations.•Compositional features plus diverse data enable authentic reconstruction.
External IDs:dblp:journals/nn/ShirakawaNTAMMK25
Loading