ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback.

Ming Li, Taojiannan Yang, Huafeng Kuang, Jie Wu, Zhaoning Wang, Xuefeng Xiao, Chen Chen

ECCV 2024   [Website] [Code] [Demo] [Slides]

• Existing efforts in controllable generation still perform poorly in controllability, with generated images deviating from input conditions.
• We show that pre-trained discriminative models can serve as powerful visual reward models to improve the controllability in a cycle-consistency manner.
• We disrupt the consistency between input images and conditions, and enable the single-step denoising for efficient reward fine-tuning.
• We provide a unified and public evaluation of controllability and demonstrate that ControlNet++ comprehensively outperforms existing methods.

AlignDet: Aligning Pre-training and Fine-tuning in Object Detection.

Ming Li^*, Jie Wu^*#, Xionghui Wang, Chen Chen^#, Jie Qin, Xuefeng Xiao, Rui Wang, Min Zheng, Xin Pan.

ICCV 2023   [Website] [Code] [Poster]

• We point out that existing detection algorithms are constrained by the data, model, and task discrepancies between pre-training and fine-tuning.
• We propose AlignDet to align these discrepancies, which constructs detection-oriented pre-training by learning classification and regression knowledge.
• AlignDet makes the first attempt to fully pre-train all kinds of detectors using a completely unsupervised paradigm, by integrating pre-trained backbones.
• AlignDet can achieve significant improvements across diverse protocols, such as detection algorithm, model backbone, data setting, and training schedule.

FreeSeg: Unified, Universal and Open-Vocabulary Image Segmentation.

Jie Qin^*, Jie Wu^*#, Pengxiang Yan, Ming Li, Ren Yuxi, Xuefeng Xiao, Yitong Wang, Rui Wang, Shilei Wen, Xin Pan, Xingang Wang^#.

CVPR 2023   [Website] [Code]

• We propose FreeSeg, a generic framework to accomplish Unified, Universal and Open-Vocabulary Image Segmentation.
• FreeSeg optimizes an all-in-one network via one-shot training and adaptive prompt learning, making a single model work for diverse segmentation tasks.
• FreeSeg establishes SOTA results on three Open-Vocabulary segmentation tasks, and outperforms the best task-specific architectures by a large margin.

DiffusionEngine: Diffusion Model is Scalable Data Engine for Object Detection.

Manlin Zhang^*, Jie Wu^*#, Yuxi Ren^*, Ming Li, Jie Qin, Xuefeng Xiao, Wei Liu, Rui Wang, Min Zheng, Andy J. Ma^#.

arXiv 2023   [Website] [Code]

• We reveal the Diffusion Model is a scalable data engine for object detection.
• We present DiffusionEngine to provide high-quality and diversity detection data, by a pre-trained diffusion model and an effective Detection-Adapter.
• We demonstrate data scaling-up via DiffusionEngine can achieve significant improvements in diverse scenarios, including various detection algorithms, self/semi-supervised pretraining, data/label scarce and cross-domain setting.