Transformer model improves image detail without raising global cost
Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers
Computer Vision and Pattern RecognitionArtificial IntelligenceMachine LearningNeural and Evolutionary Computing
Summary
Generating detailed images with Transformers usually gets slower as more parts of the image are processed globally. The authors found that large-scale patterns need global connections, but fine details can be handled locally. They created a model called SpecSlot that separates these two tasks, allowing it to add image detail without slowing down global processing. Tests show it creates high-quality images faster and more efficiently than some other methods.
What this means in practice
- •For image generation engineers: Design image generators that produce finer details with lower computational cost by separating global and local attention on pixel data.
- •For graphics software developers: Implement efficient pixel-space generative models for higher-resolution images without a large increase in processing power.
Authors
Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu
Abstract
Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. We observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally. Based on this insight, we introduce SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots interact locally. This allows the model to increase spectral detail without enlarging the global sequence. A spatial head further adds a local correction to the spectral prediction, while the generative state, target, and loss remain in pixel space. On ImageNet-256, SpecSlot achieves a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512, it improves FID by 2.84 over the control with 3.9% fewer FLOPs, and comes within only 0.69 FID of JiT-B/16 at 38.2% lower cost. On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.