Ant Technology Research Institute Interactive Intelligence Lab – 13 Papers Accepted at CVPR 2023 and Recent AI Research Highlights

CVPR 2023, one of the three top computer‑vision conferences, received 9,155 valid submissions (a 12% increase over 2022) and accepted 2,360 papers, resulting in an acceptance rate of 25.78%.

The Interactive Intelligence Lab of Ant Technology Research Institute had 13 papers selected, primarily focusing on generative models and 3D vision.

As one of the first labs established by the institute, the Interactive Intelligence Lab researches generative models, 3D vision, multimodal interaction, and human‑computer interaction. Its members come from Tsinghua University, Zhejiang University, USTC, CUHK, HKUST and collaborate with institutions such as Shanghai Jiao‑Tong University, Oxford, and UC‑Berkeley.

In the past year the lab has pursued a range of projects including basic generative models (GANs, diffusion models), controllability and interpretability of generative models, 3D‑aware generative perception, video generation, digital humans and digital scenes. To date the lab has had 21 papers accepted at top conferences: ICML 2022 (2), NeurIPS 2022 (4), TPAMI 2022 (1), ICLR 2023 (1), and CVPR 2023 (13), addressing fundamental problems in computer‑generated models and 3D vision.

02 Selected Papers

CVPR 2023 Acceptance: Learning 3D‑aware Image Synthesis with Unknown Pose Distribution . Existing 3D‑aware image synthesis methods rely on accurate 3D‑pose priors, which are costly to obtain. The lab proposes PoF3D, which equips the generator with a pose learner that infers pose from latent space and adds a pose‑prediction branch to the discriminator, eliminating the need for explicit pose priors. Experiments on multiple datasets show state‑of‑the‑art image and geometry quality without any pose prior.

ICLR 2023 Acceptance: Towards Smooth Video Composition . The paper introduces StyleSV, a new video generation approach based on generative adversarial networks. By modeling temporal relationships at short, medium, and long ranges, the method significantly improves video synthesis quality across several benchmarks, providing a simple yet effective baseline for GAN‑based video generation.

03 Ant Technology Research Institute and Its Labs

The institute aims to conduct useful and imaginative research, targeting the digital and intelligent future, advancing frontier technologies, and fostering deep industry‑academia‑research integration to strengthen China’s digital economy. Besides the Interactive Intelligence Lab, the institute hosts six other labs: Database, Graph Computing, Cryptography, Programming Languages & Compilers, and Computing Systems.

Having just begun, the institute presents the achievements of the Interactive Intelligence Lab as a one‑year report and an invitation for more collaborators to join the pursuit of technological progress.

Please receive the Ant Technology Research Institute’s report card