Bingqi Ma

I'm a researcher at the Base Model R&D Department of SenseTime Research, supervised by Guanglu Song and Yu Liu. I obtained my Bachelor's degree in 2019 and Master's degree in 2022 from School of Computer Science and Engineering at Beihang University.

Email  /  Google Scholar  /  Github

Research

Currently, my main research focuses on the pretraining and application techniques of generative models. I lead a small team responsible for the pretraining research of SenseMirage, a powerful and artistic image diffusion model. Prior to this, my research mainly centered around the applications of large-scale visual models in downstream tasks, including retrieval, recognition, and perception. We are recruiting research interns specializing in diffusion models. Welcome to email me for further details.

A plug-and-play adaptation method that allows diffusion models to be conditioned on multiple reference images alongside the textual prompt leveraging the image–text comprehension capabilities of multimodal large language models

A pioneering study on employing large language models as text encoders for diffusion models. It offers rigorous analyses and practical solutions for effectively leveraging decoder-only large language models in this role, and its methods have been adopted in prominent projects such as Hunyuan-Video.

A multimodal large language model capable of automatically selecting the vision encoder that best matches the task based on the user’s instruction, achieving advanced performance on multiple benchmarks in various tasks.

A large-scale face recognition training algorithm that is robust to noisy data and automatically resolves conflicts introduced by fine-grained noise.

A domain adaptive object detection method for multi-source data, effectively mitigating the issue of knowledge degradation in multi-source fusion scenarios.

Tech Report & Pre-print Papers

A post-training approach for diffusion models, grounded in adversarial generative training that significantly enhances both distribution alignment and image quality.

A diffusion model-based image face-swapping method that decouples identity information and attribute information, demonstrating superior identity similarity and attribute consistency.

A framework to repurpose pretrained score/flow-based generative models by reversing their generation process to extract unsupervised, hierarchy-selectable features for downstream discriminative tasks.

Top1 solution for the Unconstrained Track of ICCV 2021 MFR Challenge. The work spanned the entire pipeline, from data cleaning and model architecture design to training optimization. Portions of these techniques were contributed to Sensetime FRVT submissions.