TextAE: A Benchmark and Flow-Matching Baseline for Text-Controlled Layered Animation
Abstract
Layered animation is a common representation in motion graphics and interface design, but authoring high-quality motion remains a manual and iteration-heavy process. Existing multimodal generation systems often target pixel-space video synthesis or full animation-program generation, making them a limited match for workflows that start from an existing visual layer and require lightweight, editable motion control. In this paper, we study text-controlled layered animation with editable transformation tracks: given a static image layer, a text instruction, and an initial transformation state, the goal is to generate a temporally coherent transformation track that can be further edited in professional animation tools. To support empirical study of this setting, we present TextAE, a benchmark built from 9,585 real animation projects and 75,520 expert-designed layer-wise transformation tracks, with a text-annotated subset of 17,774 layers. TextAE is our primary contribution and provides a foundation for studying this task. We also provide Track-FM, a flow-matching baseline that jointly predicts full transformation tracks from multimodal conditions while encouraging smooth and loopable motion. Experiments indicate that TextAE provides a useful testbed for controllable layered animation generation and that Track-FM serves as a strong first baseline for producing faithful, smooth, and editable animation effects.