Your Data Manifold is Secretly a Reward Model: Shell-LCC for Text-to-Video Generation

2026-06-29

AI Summary

Existing text-to-video generation methods rely on external reward signals for alignment, suffering from high computational costs, dependence on manual annotations, and difficulty in optimizing fine-grained details. This work proposes Shell-LCC, a novel approach that models the manifold structure of high-quality supervised fine-tuning data by constraining latent representations to an isotropic shell rather than collapsing them toward the geometric center, thereby avoiding mean regression and effectively preserving high-frequency details. Shell-LCC constructs an intrinsic, dense, and differentiable reward signal through local coordinate coding (LCC), eliminating the need for additional annotations or external reward models. Experimental results demonstrate that the method significantly enhances video realism and mitigates low-level artifacts such as over-smoothing and motion blur.

Abstract

Recent text-to-video (T2V) diffusion models rely heavily on auxiliary reward signals (e.g., via reward models or DPO) to align generated content with human aesthetics and improve realism. These signals, however, incur substantial computational overhead, require costly human annotations, and often yield limited improvement in fine-grained local details. In this paper, we argue that your data manifold is secretly a reward model. By explicitly modeling the manifold structure of high-quality Supervised Fine-Tuning (SFT) data and encouraging video latents to lie on this manifold, we derive dense, differentiable, and nearly cost-free reward signals that significantly improve video quality, particularly in mitigating low-level distortions. Our modeling builds upon Local Coordinate Coding (LCC), which captures the `skeleton' of the manifold. However, directly applying LCC suffers from mean regression, pulling latents toward the geometric mean and losing high-frequency details. We therefore extend it to Shell Local Coordinate Coding (Shell-LCC), which models the manifold `surface' as an isotropic shell to align with the true high-density region. Experiments demonstrate that our approach improves realism, enhances high-frequency details, reduces over-smoothing artifacts, and alleviates motion blur.

Problem

• text-to-video generation
• reward model
• data manifold
• local details
• computational overhead

Innovation

• Shell-LCC
• data manifold
• text-to-video generation
• reward-free alignment
• Local Coordinate Coding