SpecSlot separates image detail from global-attention sequence length
A 22-page arXiv paper under review at ICLR 2027 introduces SpecSlot, a pixel-space rectified-flow Transformer designed to increase spectral detail without enlarging the global-attention sequence. It expands each spatial anchor into slots assigned to disjoint bands of DCT coefficients. Only the lowest-frequency slot participates in global attention, while all slots interact locally; a spatial head adds a local correction to the spectral prediction. The authors report comparable final FID and faster convergence versus a FLOPs-matched JiT control on ImageNet-256. On ImageNet-512, they report a 2.84 FID improvement with 3.9% fewer FLOPs, while on DPG-Bench the model reportedly exceeds JiT-B/8 by 1.06 at 40.6% lower forward cost. These results are author-reported and have not been independently verified.
The paper proposes separating low-frequency global communication from local modeling of high-frequency detail. Its reported results show that adding spectral detail through local slots can coincide with lower computational cost, although the findings still require independent evaluation.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.40647 (cs)
[Submitted on 29 Oct 2026]
Title:Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers
Authors:Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu
View a PDF of the paper titled Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers, by Jiongfan Yang and 4 other authors