Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

SpecSlot decouples output detail from global attention by splitting a fixed token budget into spatial anchors and spectral slots, sending only the lowest-radial slot per anchor through global blocks.

Pixel-space generative Transformers represent full-resolution images as patch tokens, so the patch grid simultaneously determines input reading, output detail, and the number of tokens in global attention. Finer detail therefore increases attention cost quadratically: on a 256x256 image, reducing patch size from 16x16 to 8x8 raises tokens from 256 to 1,024 and makes the dense-attention quadratic term 16x larger. Existing pixel-space methods compress or restrict communication but largely keep output detail tied to the spatial patch grid.

SpecSlot uses a shared pixel-space rectified-flow objective and replaces the patch grid with W spatial anchors, each expanded into K indexed slots, so N=WK. Read extracts non-overlapping sxs input cells, projects them to anchor vectors, applies a residual depthwise input mixer, and expands each anchor into K hidden slot states. Allocate assigns each slot a disjoint radial group of DCT coefficients and decodes a lapped-DCT spectral base plus an anchor-aligned spatial correction. Route sends only slot 0, owner of the lowest-frequency group, into global blocks, while all K slots mix locally within their anchor; global-global-local-local blocks repeat three times. With W=256 and K=4 at ImageNet-256, the global sequence holds W+C=288 tokens instead of WK+C=1,056.

• On ImageNet-256, SpecSlot W256/K4 improves over JiT-B/16 and uses 40.9% fewer FLOPs than JiT-B/8, which remains the quality leader. §4.2 ImageNet Results
• On ImageNet-512, SpecSlot W256/K4 lowers FID by 2.84 over JiT-CM with 3.9% fewer GFLOPs and trails JiT-B/16 by 0.69 FID at 38.2% lower cost. §4.2 ImageNet Results
• On text-to-image generation, SpecSlot surpasses JiT-B/8 on DPG-Bench by 1.06 at 40.6% lower forward cost, reporting 62.21 versus 61.15. §4.3 Text-to-Image Generation
• At fixed W256/K4/B1536, GGLL reaches FID 6.59 versus 6.81 for the FLOPs-matched Dense-global-CM control, and saves 42.8% of the full-global cost. §5.1.3 Carrier Routing
• Changing the global carrier from slot 0 to slot 1, 2, or 3 raises FID from 6.59 to 6.91, 7.21, or 6.97. §6.2 Carrier Identity and Routed Communication
• At fixed W=256, the entropy-effective slot fraction falls from 0.803 at K=2 to 0.495 at K=8, and the last K doubling buys only 0.31 FID. §5.1.1 Anchor Density and Coefficient Subdivision

• On ImageNet-512, the two operating points are consistent with within-anchor allocation but do not by themselves isolate the cause of the larger margin over JiT-CM (stated in §4.2).
• The trend favoring more spatial anchors over more K is measured only within the tested N=2,048 configurations; the paper explicitly does not claim it generalizes beyond that range (stated in §5.1.1).
• Additional spectral slots show diminishing returns: the entropy-effective slot fraction declines from 0.803 at K=2 to 0.495 at K=8, and the paper does not promise each added slot contributes equal information (stated in §5.1.1 and §E.2).
• The text-to-image table reports GenEval 26.99 for SpecSlot versus 29.78 for JiT-B/8, and the paper does not analyze this lower GenEval result; text-to-image evaluations are at 256x256 (reported in §4.3).

• Token count splits as N=WK: W sets anchor density and K sets spectral subdivision, so the global sequence can stay at W carriers. §3.2 SpecSlot Architecture
• Only slot 0, assigned the lowest radial group, enters global blocks; all K slots mix within each anchor under a GGLL schedule repeated three times. §3.2.3 Route: Carrier and Local Attention
• On ImageNet-256 with s=16 and K=4, the global sequence holds W+C=288 tokens instead of WK+C=1,056. §3.2.3 Route: Carrier and Local Attention
• The global quadratic cost grows with W, not WK: the GGLL quadratic term is 6(W+C)^2 d + 6WK^2 d versus 12(WK+C)^2 d for full-global. §3.2.3 Route: Carrier and Local Attention
• Within tested N=2,048 configurations, shifting the split toward more anchors lowers FID monotonically, while increasing K has diminishing returns; the paper does not claim this rule generalizes beyond the measured range. §5.1.1 Anchor Density and Coefficient Subdivision
• At the same spectral output, spatial input surpasses frequency input, and the dual-basis hybrid reaches FID 6.59, 0.40 below the best single-basis variant at 6.99. §5.1.2 Input and Output Representations
• GGLL is better and faster than the FLOPs-matched Dense-global-CM control; full-global attention reaches lower FID but at 74.8% higher cost and lower throughput. §5.1.3 Carrier Routing
• The lowest-radial carrier assignment matters: replacing slot 0 with another slot raises mean FID in every case, and this band crosses DNR=1 earliest. §6.2 Carrier Identity and Routed Communication