Role-Coupled Tokenization Reads, Allocates, and Routes Tokens for Pixel-Space Generative Transformers
SpecSlot assigns fixed spectral roles to pixel-space image tokens, cutting compute and improving DPG-Bench with local slots and global attention.
Key Takeaways
- SpecSlot is a role-coupled tokenization method for pixel-space generative Transformers that assigns fixed spectral roles to image tokens.
- It separates visual processing into low-frequency carriers for global attention and local slots for local interactions, reducing unnecessary global compute.
- On ImageNet-512, SpecSlot-K4 uses 38.2% fewer FLOPs than JiT-B/16.
- For 256×256 text-to-image generation, it cuts forward compute from 218.77 to 129.86 GFLOPs while improving DPG-Bench from 61.15 to 62.21.
- Paper, project page, and code are available, highlighting a lower-compute direction for pixel-space image generation.
Meta Description: SpecSlot cuts compute in pixel-space generative Transformers, reducing FLOPs and improving DPG-Bench on text-to-image tasks using fixed spectral token roles.
Tags: SpecSlot, pixel-space Transformers, generative AI, DPG-Bench, FLOPs
Featured Image: Abstract friendly illustration of image tokens split into global low-frequency carriers and local slots, no real people, no logos, no text.
What SpecSlot changes
SpecSlot is a role-coupled tokenization method for pixel-space generative Transformers. It separates fine visual detail from global communication by assigning tokens fixed spectral output roles: low-frequency carriers perform global attention, while the remaining slots interact locally.
Reported compute and quality
On ImageNet-512, SpecSlot-K4 uses 38.2% fewer FLOPs than JiT-B/16. For 256×256 text-to-image generation, it reduces forward compute from 218.77 to 129.86 GFLOPs while improving DPG-Bench from 61.15 to 62.21.
Links and what to watch
The paper, project page, and code are listed at https://arxiv.org/abs/2610.40647, https://yangjiongfan.site/specslot/, and https://github.com/JiongfanYang/SpecSlot. The reported numbers show a lower-compute direction for pixel-space image generation, while the released resources let readers inspect the method.

Comments (0)