Tag
The paper introduces MilliVid, a method for improving long-range consistency in video generation by using a multi-scale autoencoder to compress frames into hierarchical tokens and then generating them with a coarse-to-fine diffusion model, outperforming baselines on Minecraft videos.
This paper proposes CAT, a cross-scale aligned transformer that enforces consistency between intermediate and final GAN outputs to resolve trajectory misalignment, achieving state-of-the-art FID of 1.56 on ImageNet-256.
This paper proposes a nested spatiotemporal forecasting framework that uses spectral clustering to construct semantically coherent macro-level regions, which provide top-down guidance for fine-grained micro-level predictions. Experiments on high-dimensional datasets show consistent improvements over state-of-the-art baselines.
HL-OutPaint is a coarse-to-fine video outpainting framework for high-resolution long-range videos, using global coarse guidance to enable large spatial extrapolation while maintaining spatio-temporal consistency.
This paper investigates how 1D coarse-to-fine token structures in autoregressive models improve test-time search efficiency compared to classical 2D grid tokenization. The authors show that such ordered tokens enable better test-time scaling and even training-free text-to-image generation when guided by image-text verifiers.