SpecSlot separates global image structure from local detail

What the study found

In an arXiv preprint, the authors introduce SpecSlot, a pixel-space rectified-flow Transformer. Instead of treating each image patch as a single token, it expands each spatial anchor into slots assigned separate bands of discrete cosine transform (DCT) coefficients. Only each anchor’s lowest-frequency slot enters global attention; all slots interact locally. A spatial head adds a local correction, while the generative state, target and loss remain in pixel space.

The authors report that on ImageNet-256, SpecSlot reached a comparable final Fréchet inception distance (FID) to a JiT control matched for floating-point operations (FLOPs), while converging faster. On ImageNet-512, they report a 2.84-point FID improvement over the control with 3.9% fewer FLOPs. They also report an FID within 0.69 of JiT-B/16 at 38.2% lower cost. For text-to-image generation, SpecSlot exceeded JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost.

Why it matters

Global attention becomes more expensive as its sequence grows. The proposed routing scheme allows finer spectral subdivision without adding entries to that global sequence, potentially offering a more efficient way to model image detail.

Limits and unknowns

These are author-reported preprint results, not established general performance guarantees. The abstract does not provide full training and evaluation settings, absolute benchmark scores, uncertainty estimates or details of how costs were measured. It also does not establish how the approach performs across other datasets or model scales.