NUS Proposes Wiener-Informed Decoding to Speed Up Pixel-Space Flow Model Training
NationalUniversityofSingapore · hf · 2026-10-29
Researchers at NUS HPC-AI Lab (Ziqiao Wang & Yang You) introduce Wiener-informed Decoding (WiD), tackling the difficulty of learning velocity fields directly in pixel space.
- Key observation: frozen Wiener estimates (coarse second-order image statistics) improve velocity learning, with gains extending beyond linear predictions; frozen bases and pixel decoding bring complementary gains.
- Method: WiD uses statistical knowledge both for direct prediction and to organize decoder computation — a trainable Wiener prediction path with coefficient interfaces injects features into the native decoder and returns coefficient corrections, all trained jointly under the original pixel-space velocity objective.
- Results: across model scales and decoder architectures, WiD accelerates training and improves generation quality with modest added parameters.
More from Multimodal
- Xmax ships X2.1 real-time video editing at 1080P for $0.0016/sec, live face and outfit swaps — Scobleizer · 2026-10-29
- Genie 4 demo: chilling in a Japanese onsen with pettable capybaras — venturetwins · 2026-10-29
- Mirage launches Tesseract: turn any video into an agent-editable project — gizakdag · 2026-10-29
- PixVerse-Generated Cat Working Out on Playground Equipment Is Uncannily Convincing — Additional-Engine402 · 2026-10-29
- Seedance 2.5 prompt shared: one long prompt generates a Pixar-style family animation — aftahi_ai · 2026-10-29
- AI music video: Claudia's first day of robot school — ClaudiaSlop · 2026-10-29