Two 360 AI Research Projects Accepted to ICML 2026
Research built around precision and control
Two projects from 360 AI Research have been accepted to ICML 2026: RevealLayer and FG-CLIP 2. Both address a limitation of broad multimodal capability—models must understand fine differences and produce results that people can control.
RevealLayer: editable decomposition from one image
RevealLayer separates a single RGB image into independent RGBA layers. Region-aware attention, occlusion-guided modeling, and alpha-aware training help the model isolate selected objects, reconstruct hidden content, and restore the background. The result turns a flattened image into material that can be edited and recomposed.
FG-CLIP 2: fine-grained bilingual alignment
FG-CLIP 2 improves detailed image–text alignment while retaining an efficient dual-encoder design. It uses stronger Chinese data, dynamic-resolution training, and a redesigned objective to support fine-grained recognition, retrieval, open-vocabulary detection, and region-level understanding. The project is evaluated across 29 tasks in eight English and Chinese categories.
Together, these projects connect research advances with practical products for creative editing and multimodal retrieval.