Paper Overview
Research area: Computer Vision Authors: Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu arXiv: 2610.40647
Abstract (English translation)
Pixel-space generative Transformers typically use one token per image patch, so capturing finer detail requires more tokens and makes global attention increasingly expensive. The authors observe that finer detail does not necessarily require more global communication: low-frequency structure benefits from global exchange, while high-frequency detail can be modeled locally.
Based on this insight, the paper introduces SpecSlot, a pixel-space rectified-flow Transformer that decouples output detail from the length of the global attention sequence. SpecSlot expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients; this separates anchor density from spectral subdivision. Only the lowest-frequency slot from each anchor participates in global attention, while all slots participate in local modeling.
Related Discussion (from the post's Chinese summary)
The post also includes a discussion of micro-edits in film editing:
- Macro, content-level cuts shape a film's high-level narrative, while subtle local cuts control pacing and guide visual attention — e.g., adjusting the exact timing of a segment to control the length of a meaningful pause, or cutting to a shot highlighting a character's reaction.
- Such fine-grained adjustments are slow and cumbersome in standard editing tools, since users must map high-level narrative goals to low-level, frame-by-frame changes.
- The described work conducts a formative study with professional editors to understand the structure of micro-edits, derives a mathematical formalization of that structure, and builds tools that leverage AI and optimization to make exploring and achieving micro-edits easier. The tools are validated through multiple user studies and extensive video examples.
*Auto-collected on 2026-10-31.*