Mode Seeking meets Mean Seeking for Fast Long Video Generation

Scaling video generation from seconds to minutes faces a critical bottleneck: while short-video data is abundant and high-fidelity, coherent long-form data is scarce and limited to narrow domains. While multi-resolution image training works because higher resolution is largely an interpolation of the same underlying patch distribution, training across video lengths is fundamentally different: a longer video is an extrapolation that must invent new events and causal structure beyond the short-clip horizon. To address this, we propose a training paradigm where Mode Seeking meets Mean Seeking, decoupling local fidelity from long-term coherence from a unified representation via a Decoupled Diffusion Transformer. Our approach utilizes a global Flow Matching head trained via supervised learning on long videos to capture narrative structure, while simultaneously employing a local Distribution Matching head that aligns sliding windows to a frozen short-video teacher via a mode-seeking reverse-KL divergence. This strategy enables the synthesis of minute-scale videos that learns long-range coherence and motions from limited long videos via supervised flow matching, while inheriting local realism by aligning every sliding-window segment of the student to a frozen short-video teacher.

Today’s video AI systems can create impressive short clips, but making a coherent minute-long video is much harder: the model must keep track of characters, scenes, and events while still rendering each moment sharply. A common way to train long-video models helps them follow the overall story, but it can make fine details look averaged or blurry. Another way uses a strong short-video model to improve local realism, but by itself it does not understand what should happen over a longer time. Our work combines these two strengths. We train one part of the model to learn long-range structure from real long videos, and another part to borrow visual quality and motion realism from a short-video expert. By separating these two goals instead of forcing one component to satisfy both, our model can generate longer videos that are both more coherent and more visually detailed. This could make video generation more useful for storytelling, simulation, and creative tools where users need more than a few seconds of realistic motion, a critical part towards world modeling.