Creating 3D objects from images typically starts with generating a rough 3D skeleton. Current methods struggle to control the skeleton's orientation and often produce structures that miss visible parts or add nonexistent ones.
Coarse3D addresses both issues. It ties the 3D coordinate system to the first input image, so orientation is defined by the view rather than a guessed front. To improve faithfulness, it trains a reconstruction task alongside generation and uses an attention mechanism that forces each 3D point to attend to the correct image pixels.
Coarse3D improves the coarse skeleton quality, raising an accuracy measure from 0.54 to 0.86 on a benchmark. When used as a drop-in replacement, it boosts final 3D generation without retraining, making image-to-3D more reliable.