Resolution-Stable Arbitrary Style Transfer via Cross-Scale Low-Frequency Attention Anchoring

Resolution-Stable Arbitrary Style Transfer via Cross-Scale Low-Frequency Attention Anchoring

Abstract—Changes in content resolution can cause StyTR-2 to shift its attention correspondences and output texture organization. To address this problem, we propose a training-free cross-scale low-frequency attention-response anchoring method. A frozen low-resolution branch provides a reference whose decoder responses are spatially aligned with those of the high-resolution branch. Only the low-frequency component of the cross-scale response difference is injected, correcting global correspondences while retaining high-resolution texture details. Mutually exclusive diagnostic, development, and test sets are used to evaluate consistency in attention responses, perceptual features, and pixel outputs across three resolution pairs. The proposed method consistently reduces cross-resolution drift. Ablations over frequency range, intervention location, and decoder stage further support the roles of low-frequency constraints, attention-space intervention, and multi-stage anchoring. Overall style statistics remain stable, although the method incurs a small loss of local structural fidelity and additional inference latency. Because no model weights are updated, the method can serve as a resolution-stability extension to StyTR-2 for preview-to-high-resolution export workflows.

Keywords- Arbitrary Style Transfer; Stytr-2; Cross-Scale Attention; Low-Frequency Anchoring; Resolution Stability