Dress-1-to-3: Single Image to Simulation-Ready 3D Outfit (dress-1-to-3.github.io)

Researchers presented Dress-1-to-3, a pipeline that converts a single in-the-wild photo of a clothed person into separated, simulation-ready 3D garments (with sewing patterns) and a human avatar. The system first predicts a coarse sewing pattern from the image, then uses a pre-trained multi-view diffusion model to synthesize orbital views that act as pseudo ground-truth for 3D shape and pose. A differentiable garment simulator “sews” and drapes the pattern onto the posed human and optimizes shape, material parameters, and geometric regularizers to align with the generated multi-view images. The pipeline also adds texture generation and human motion modules to produce animated, physics-plausible garment demonstrations. This work is significant because it moves beyond monolithic, fused reconstructions to output separable, physically plausible garments that are ready for cloth simulation—directly addressing needs in virtual try-on, AR/VR, and animation. Technically notable is the fusion of generative diffusion priors (to hallucinate multi-view cues from one image) with differentiable physics-based simulation and pattern optimization, enabling recovery of both geometric rest shapes and simulation parameters. Using synthetic multi-view imagery as supervision is a pragmatic solution to single-view ambiguity and suggests a new direction: combining learned priors with differentiable physical models to produce actionable 3D assets for downstream ML and graphics applications.