With advancements of diffusion models in generating high-quality, diverse, and creative content, subject-driven image editing based on diffusion models has been growing attention. However, it is challenging to perform precise control in image editing as minor changes in the input can lead to significant differences in generated concepts and content. Prevalent methods face significant limitations in terms of large numbers of trainable parameters and inference latency. Hence, we propose an efficient tensor-based fine-tuning method, Tucker/TT-based Fine-Tuning (TuckerFT and TTFT), for subject-driven image editing on diffusion models. Specifically, we first provide Tucker/TT-based Fine-tuning Convolutional Adapter and Attention Adapter to achieve parameter-efficient fine-tuning in convolutional and attention layers of diffusion models. Additionally, we further present a more sound initializing strategy for the Tucker/TT-based Adapters to guarantee steady training. Extensive experimental results demonstrate that TuckerFT/TTFT surpasses state-of-the-art methods on subject similarity, trainable parameters, and training duration.

As the visual interpretations for convolutional neural networks (CNNs), backpropagation attribution methods have been garnering growing attention. Nevertheless, majority of these methods merely concentrate on the ultimate convolutional layer, leading to tiny and concentrated interpretations that fail to adequately clarify the model-central attention. Therefore, we propose a precise attribution method (i.e., Holistic-CAM) for high-definition visual interpretation in the holistic stage of CNNs. Specifically, we first present weighted positive gradients to guarantee the sanity of interpretations in shallow layers and leverage multi-scale fusion to improve the resolution across the holistic stage. Then, we further propose a denoising strategy based on the fundamental scale component to eliminate the faithless attribution derived from fusing larger-scale features. The proposed method is capable of simultaneously rendering fine-grained and faithful interpretations for CNNs from shallow to deep layers. Extensive experimental results demonstrate that Holistic-CAM outperforms state-of-the-art methods on common-used benchmarks, including deletion and insertion, energy-based point game as well as Remove and Debias on ImageNet-1k, it also passes the sanity check easily.

Video dubbing is a new type of language learning practice. Because of the fun it brings into learning, video dubbing mobile applications have become quite popular. During video dubbing, learners not only mimic characters' pronunciations but also other voicing characteristics, e.g., emotions. In this study, we collected a corpus from Liulishuo English learning App's video dubbing practices and investigated three research questions. We find that L2 learners do try to mimic emotions via their voices, and human raters tend to focus more on the emotion production aspect when evaluating dubbing performance. Furthermore, we develop state-of-the-art speech emotion recognition technology and show it can be used for evaluating emotion production automatically.