Presentation Information
[2K6-GS-7d-05]Frequency-ordered Autoregressive Image Generation Model with Adaptive Scheduling
〇Daiki Miyake1, Masahiro Suzuki1, Yutaka Matsuo1 (1. The University of Tokyo)
Keywords:
Image generation,Autoregressive model
Recent advances in LLMs and VLMs have stimulated autoregressive image generation based on next-token prediction over discretized image patches. Previous work shows that generating patches in frequency order, from low to high resolutions, improves image quality compared with raster-scan ordering.
However, since it relies on heuristically predefined resolution scheduling, it is difficult to discover optimal schedules for other training data.
We propose a continuous-time extension of frequency-ordered autoregressive image generation that enables flexible scheduling at inference time.
Additionally, we propose a sampling algorithm that leverages the past prediction to mitigate error accumulation.
Experiments on CIFAR-10 demonstrate that adaptive scheduling achieves higher generation quality than models with fixed frequency orders, improving FID by more than 50\%.
Furthermore, we show that even varying the number of steps by a single step leads to a noticeable trade-off in generation quality.
However, since it relies on heuristically predefined resolution scheduling, it is difficult to discover optimal schedules for other training data.
We propose a continuous-time extension of frequency-ordered autoregressive image generation that enables flexible scheduling at inference time.
Additionally, we propose a sampling algorithm that leverages the past prediction to mitigate error accumulation.
Experiments on CIFAR-10 demonstrate that adaptive scheduling achieves higher generation quality than models with fixed frequency orders, improving FID by more than 50\%.
Furthermore, we show that even varying the number of steps by a single step leads to a noticeable trade-off in generation quality.