SpecSlot Boosts Pixel-Space Image Generation: +2.84 FID on ImageNet-512 at Lower Cost
Jiongfan Yang · hf · 2026-10-30
Researchers from Peking University, NTU and SenseTime introduce SpecSlot, a pixel-space rectified-flow transformer that decouples output detail from global attention length. Each spatial anchor is expanded into spectral slots with disjoint DCT bands; only the lowest-frequency slot joins global attention while all slots interact locally, adding spectral detail without a longer sequence. On ImageNet-512 it improves FID by 2.84 over a FLOPs-matched control with 3.9% fewer FLOPs, sits within 0.69 FID of JiT-B/16 at 38.2% lower cost, and beats JiT-B/8 on DPG-Bench by 1.06 for text-to-image at 40.6% lower forward cost.
More from Multimodal
- AI-generated Winamp lighter concept goes viral with retro EQ and playlist details — Delahuntagram · 2026-10-30
- Rapid-H3 generates video + audio in ~2 minutes on a single DGX Spark — vchauhan_ · 2026-10-30
- Midjourney Style Code sref 3594718371 Produces Striking Metallic Armor Fashion Portraits — sergeantsref · 2026-10-30
- Native 4K with MoSS Image: RoPE NTK vs Upscale Workflows Put to the Test — No-Sleep-4069 · 2026-10-30
- Open-source openjutsu adds 17 relight presets and two-pass Qwen3-VL scene describing — Devajyoti1231 · 2026-10-30
- PhysInsert from Beihang and Meitu Achieves Physically Plausible Reference-Based Object Insertion — Yue Gong · 2026-10-30