ATATA: One Algorithm to Align Them All

TL;DR

ATATA introduces a universal, inference-based method for structurally aligned generation across images, videos, and 3D objects by performing joint segment transport within a structured latent space using Rectified Flow. By transporting a line segment between two samples along the velocity field $v_{\Theta}$ and incorporating a smoothness regularization via an anchor velocity, the approach achieves high structural alignment with fast inference compared to SDS-based rivals. The method is backboned by three state-of-the-art rectified-flow models operating on voxels, pixels, and video frames, and is validated across three modalities with metrics capturing alignment, prompt fidelity, and perceptual quality, often surpassing editing-based and joint-generation baselines. The resulting framework enables scalable, multi-modal aligned generation and synthetic data creation, with substantial speedups in 3D pipelines and strong performance in images and videos.

Abstract

We suggest a new multi-modal algorithm for joint inference of paired structurally aligned samples with Rectified Flow models. While some existing methods propose a codependent generation process, they do not view the problem of joint generation from a structural alignment perspective. Recent work uses Score Distillation Sampling to generate aligned 3D models, but SDS is known to be time-consuming, prone to mode collapse, and often provides cartoonish results. By contrast, our suggested approach relies on the joint transport of a segment in the sample space, yielding faster computation at inference time. Our approach can be built on top of an arbitrary Rectified Flow model operating on the structured latent space. We show the applicability of our method to the domains of image, video, and 3D shape generation using state-of-the-art baselines and evaluate it against both editing-based and joint inference-based competing approaches. We demonstrate a high degree of structural alignment for the sample pairs obtained with our method and a high visual quality of the samples. Our method improves the state-of-the-art for image and video generation pipelines. For 3D generation, it is able to show comparable quality while working orders of magnitude faster.

Figures (8)

• Figure 1: Visualization of generated images, videos, and 3D shapes using our method. The left pair is (horse animal, horse skeleton), the right pair is (gothic temple, hindu temple).
• Figure 2: Method
• Figure 3: Visualization of geometry preservation between two generated images. For each example, two images are blended into one with a blending coefficient $\alpha(column)$ that depends on the column index, increasing from 0 (left side) to 1 (right side). With such blending, we show that not only geometry is preserved, but also that smooth transitions between two generations are enabled.
• Figure 4: 3D Alignment Samples
• Figure 5: Visualization of geometry preservation between source and target video frames. Larger white areas means higher difference between corresponding pixels.