A text-to-3D generation approach (TriFlow) that repurposes a pretrained feed-forward image-to-triplane reconstruction model as a frozen latent encoder, avoiding separate encoder training. Because reconstruction latents are unstructured, it adds a feature-standardization protocol, spatial weighting to emphasize important triplane regions, and a 2D perceptual rendering loss, then trains a multi-stream transformer-based rectified-flow model over the triplane tokens for linear-scaling, text-conditioned 3D generation.
This paper (TriFlow) shows that a pretrained feed-forward image-to-3D reconstruction model can be reused as a frozen latent encoder for training a 3D generative model, bridging the reconstruction and…