THREE-BODY SCATTERING FOR GENERATIVE MODELING

Modern generative models typically rely on complex, multi-step processes like diffusion, adversarial training, or autoregressive decoding. This paper asks whether it is possible to achieve high-dimensional, one-step generation using a simpler, direct distribution-matching objective that avoids the computational overhead of teacher queries or minibatch-wide all-pairs comparisons.

The authors introduce Three-Body Scattering Modeling (TBSM), a framework that treats generation as a particle-interaction problem. In this model, a generated 'projectile' is attracted to a real data source and simultaneously repelled from an independently generated source. This triplet interaction creates a constant-size, sample-level loss. To stabilize the training, the authors employ an online 'tracker' that approximates the conditional expectation of these scattering vectors, effectively smoothing the motion toward the target distribution. The generator is trained by regressing its output toward these detached, tracked targets.

TBSM successfully trains one-step generators on ImageNet-256, achieving an FID of 2.23 with pixel-space models and 1.63 with latent-space models at a single function evaluation (NFE=1). The authors provide a design map that links their scattering approach to existing paradigms like diffusion-based supervision and GAN-like dynamics. By using frozen image features, the method effectively scales to high-dimensional data without requiring classifier-free guidance or complex multi-step sampling pipelines.

This work establishes a new, efficient route for one-step generation that is theoretically grounded in energy-distance minimization. By reducing the complexity of the training objective to a simple per-projectile interaction, TBSM offers a scalable alternative to current state-of-the-art generative models, potentially simplifying the deployment of high-quality, real-time image generation systems.