CVPR2025

Composing Parts for Expressive Object Generation

Harsh Rangwani, Aishwarya Agarwal, Kuldeep Kulkarni, R. Venkatesh Babu, Srikrishna Karanam

Abstract

a full body portrait photo of a person wearing a dress with long hair and glasses PartComposer: hair, glasses, dress A photo of a yellow warbler PartComposer: beak → a hens beak a photo of a white swan, 8k, full hd PartComposer: crown → a crown of a peacock A young woman sits at a table in a beautiful, lush garden, painting by claude monet PartComposer: dress → a dress in Ukiyo-e style Figure 1. Base Generation (left) Comparison with Methods for Generations with Parts Details (right). PartComposer allows the generation of object images with specified attributes (color, style etc.) of parts for the chosen object in the base text prompt. StableDiffusion and Rich-Text [18] methods with part details either ignore the part instructions or generate inconsistent ojects.