Twitxy Saved

Shumpei Takezaki

@shumpeiMaxwell

画像生成に興味がある九大D3 / バイオ&医療AIにも興味あり/ 視覚若手の会LENS (@lens_kousiki)
PhD student of Kyushu University / Medical imaging, Generative Model
2.1K Following 1.2K Followers 3.7K Posts

Latest posts 5

Shumpei Takezaki @shumpeiMaxwell
JVT: Just Video Transformer Makes Pixel-Space Video Diffusion Simple

TL;DR: A simple Video Diffusion Transformer generates videos directly in pixel space, without VAEs, U-Nets, or cascades. Extends JiT to video with efficient spatiotemporal patching and register guidance, enabling joint RGB and geometry generation.
arxiv.org/abs/2611.18549
arxiv.org JVT: Just Video Transformer makes pixel-space video diffusion simple Direct diffusion modeling of high-dimensional raw video data is challenging, and existing pixel-space methods often rely on specialized architectures. We introduce JVT (Just Video Transformer), a...
0 replies 9 reposts 39 likes 2.1K views View on X
Shumpei Takezaki @shumpeiMaxwell
Learning Anisotropic Prediction Targets for Flow Models

TL;DR: Learns direction-dependent prediction targets for Flow Matching instead of fixing data, noise, or velocity prediction. Adapts to data variance and improves ImageNet 256×256 FID from 1.90 to 1.55.
arxiv.org/abs/2611.17767
arxiv.org Learning Anisotropic Prediction Targets for Flow Models Flow-based generative models are trained on a hand-chosen prediction target, such as the velocity, the noise, or the clean data. This choice strongly affects training and sample quality. We...
0 replies 0 reposts 18 likes 907 views View on X
Shumpei Takezaki @shumpeiMaxwell
Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between Initialization and Training

TL;DR: Heavy-tailed noise accelerates convergence to the prior but makes reverse-process learning harder. The resulting estimation errors can outweigh initialization benefits, showing that heavy-tailed noise does not necessarily improve generation or rare-event modeling.
arxiv.org/abs/2605.13175
arxiv.org Do Heavy Tails Help Diffusion? On the Subtle Trade-off Between... Recent works have proposed incorporating heavy-tailed (HT) noise into diffusion- and flow-based generative models, with the goals of better recovering the tails of target distributions and...
0 replies 5 reposts 30 likes 1.8K views View on X
Shumpei Takezaki @shumpeiMaxwell
Understanding and Learning Prediction Targets in Diffusion Models

TL;DR: なぜ高次元空間ではx-predictionがノイズ・速度予測より優れるのか?データの幾何学的構造と最適な予測対象の関係を理論的に解明.さらに,予測対象を自動学習するk-Diffを提案し,潜在空間とピクセル空間で有効性を実証.
arxiv.org/abs/2601.21419
arxiv.org Understanding and Learning Prediction Targets in Diffusion Models Recent advances in diffusion and flow matching models have highlighted a shift in the preferred prediction target -- moving from noise ($\varepsilon$) and velocity (v) to direct data (x)...
1 replies 9 reposts 62 likes 3.5K views View on X
Shumpei Takezaki @shumpeiMaxwell
Dataset Distillation via Drifting

TL;DR: A new dataset distillation approach that matches data distributions rather than optimizing classification losses. By combining attraction to real data and repulsion among synthetic samples in pretrained feature spaces, it improves diversity and achieves 27.0% ImageNet-1K accuracy with just 10 images/class.
arxiv.org/abs/2611.16642
arxiv.org Dataset Distillation via Drifting The task of dataset distillation concerns the learning of a small set of synthetic samples that can serve as an effective replacement for a large real dataset when used for downstream tasks. While...