SpecSlot
A pixel-space rectified-flow Transformer that separates spatial anchors from spectral subdivision through DCT band decoupling. Only the lowest-frequency slots participate in global attention, while high-frequency details are handled by local interaction, thereby improving generation detail while controlling global attention overhead.
Timeline (last 90 days)
SpecSlot is a pixel-space rectified-flow Transformer that expands each spatial anchor into multiple slots, with each slot responsible for a set of non-overlapping DCT coefficient frequency bands, thereby decoupling anchor density from spectral detail.
In SpecSlot, only the lowest-frequency slot of each anchor participates in global attention, while all slots interact locally, so increasing spectral detail (the number of slots) does not cause the global attention sequence length to expand.
SpecSlot introduces a spatial head that adds a local correction on top of spectral prediction, so that the generated state, objective and loss function always remain in pixel space.
On ImageNet-256, SpecSlot's final FID is comparable to the FLOPs-matched JiT control group, but it converges faster.
On ImageNet-512, SpecSlot improves FID by 2.84 relative to the control group while reducing FLOPs by 3.9%.
On ImageNet-512, the FID gap between SpecSlot and JiT-B/16 is only 0.69, while the computational cost is 38.2% lower.
On text-to-image generation tasks, SpecSlot surpasses JiT-B/8 on DPG-Bench with 40.6% lower forward cost, leading by 1.06.
The paper's authors observe that finer details do not necessarily require more global communication; low-frequency structure benefits from global information exchange, while high-frequency details can be modeled locally.
All Facts (8)
SpecSlot is a pixel-space rectified-flow Transformer that expands each spatial anchor into multiple slots, with each slot responsible for a set of non-overlapping DCT coefficient frequency bands, thereby decoupling anchor density from spectral detail.
50%UnverifiedIn SpecSlot, only the lowest-frequency slot of each anchor participates in global attention, while all slots interact locally, so increasing spectral detail (the number of slots) does not cause the global attention sequence length to expand.
50%UnverifiedSpecSlot introduces a spatial head that adds a local correction on top of spectral prediction, so that the generated state, objective and loss function always remain in pixel space.
50%UnverifiedOn ImageNet-256, SpecSlot's final FID is comparable to the FLOPs-matched JiT control group, but it converges faster.
50%UnverifiedOn ImageNet-512, SpecSlot improves FID by 2.84 relative to the control group while reducing FLOPs by 3.9%.
50%UnverifiedOn ImageNet-512, the FID gap between SpecSlot and JiT-B/16 is only 0.69, while the computational cost is 38.2% lower.
50%UnverifiedOn text-to-image generation tasks, SpecSlot surpasses JiT-B/8 on DPG-Bench with 40.6% lower forward cost, leading by 1.06.
50%UnverifiedThe paper's authors observe that finer details do not necessarily require more global communication; low-frequency structure benefits from global information exchange, while high-frequency details can be modeled locally.
50%