Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers
Paper · https://arxiv.org/abs/2610.40647
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.
First seen by Oriel on .
AI summary
SpecSlot is a pixel-space rectified-flow Transformer that increases image detail through local spectral slots while keeping the global attention sequence compact.
- Each spatial anchor expands into slots assigned disjoint DCT coefficient bands, separating anchor density from spectral detail.
- Only each anchor's lowest-frequency slot uses global attention; all slots interact locally, and a spatial head adds a local correction.
- On ImageNet-256, SpecSlot reaches comparable final FID to a FLOPs-matched JiT control while converging faster.
- On ImageNet-512, it improves FID by 2.84 with 3.9% fewer FLOPs, and on text-to-image it beats JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.
#transformers #image generation #tokenization #computer vision
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hugging Face Papers: 27, latest observation ; +4 over observed period since
Comments
Comment count reported by the source.
- Hugging Face Papers: 6, latest observation ; +2 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Papers | 2610.40647 | https://arxiv.org/abs/2610.40647 Read discussion |