DreamFusion

Instead of training on 3D datasets, DreamFusion “asks” a pretrained text-to-image model what the object should look like from many angles and uses that feedback to refine a neural field until renderings match the prompt. The final radiance field can be meshed and textured for use in engines or design tools, capturing shape and appearance with surprising fidelity from words alone. It’s particularly useful for quickly exploring character and prop ideas, generating turntables, and seeding asset pipelines for games, AR/VR, and product visualization—no paired 3D supervision required.