Peng Zhang

Aohan Zeng, Xiao Liu, Zhengxiao Du et al. · tsinghua

We introduce GLM-130B, a bilingual (English and Chinese) pre-trained language model with 130 billion parameters. It is an attempt to open-source a 100B-scale model at least as good as GPT-3 (davinci) and unveil how models of such a scale can be successfully pre-trained. Over the course of this effort, we face numerous unexpected technical and engineering challenges, particularly on loss spikes and divergence. In this paper, we introduce the training process of GLM-130B including its design choices, training strategies for both efficiency and stability, and engineering efforts. The resultant GLM-130B model offers significant outperformance over GPT-3 175B (davinci) on a wide range of popular English benchmarks while the performance advantage is not observed in OPT-175B and BLOOM-176B. It also consistently and significantly outperforms ERNIE TITAN 3.0 260B -- the largest Chinese language model -- across related benchmarks. Finally, we leverage a unique scaling property of GLM-130B to reach INT4 quantization without post training, with almost no performance loss, making it the first among 100B-scale models and more importantly, allowing its effective inference on 4$\times$RTX 3090 (24G) or 8$\times$RTX 2080 Ti (11G) GPUs, the most affordable GPUs required for using 100B-scale models. The GLM-130B model weights are publicly accessible and its code, training logs, related toolkit, and lessons learned are open-sourced at \url{https://github.com/THUDM/GLM-130B/}.

Li Hu, Xin Gao, Peng Zhang et al.

Character Animation aims to generating character videos from still images through driving signals. Currently, diffusion models have become the mainstream in visual generation research, owing to their robust generative capabilities. However, challenges persist in the realm of image-to-video, especially in character animation, where temporally maintaining consistency with detailed information from character remains a formidable problem. In this paper, we leverage the power of diffusion models and propose a novel framework tailored for character animation. To preserve consistency of intricate appearance features from reference image, we design ReferenceNet to merge detail features via spatial attention. To ensure controllability and continuity, we introduce an efficient pose guider to direct character's movements and employ an effective temporal modeling approach to ensure smooth inter-frame transitions between video frames. By expanding the training data, our approach can animate arbitrary characters, yielding superior results in character animation compared to other image-to-video methods. Furthermore, we evaluate our method on benchmarks for fashion video and human dance synthesis, achieving state-of-the-art results.

Bingchuan Li, Tianxiang Ma, Peng Zhang et al.

The StyleGAN family succeed in high-fidelity image generation and allow for flexible and plausible editing of generated images by manipulating the semantic-rich latent style space.However, projecting a real image into its latent space encounters an inherent trade-off between inversion quality and editability. Existing encoder-based or optimization-based StyleGAN inversion methods attempt to mitigate the trade-off but see limited performance. To fundamentally resolve this problem, we propose a novel two-phase framework by designating two separate networks to tackle editing and reconstruction respectively, instead of balancing the two. Specifically, in Phase I, a W-space-oriented StyleGAN inversion network is trained and used to perform image inversion and editing, which assures the editability but sacrifices reconstruction quality. In Phase II, a carefully designed rectifying network is utilized to rectify the inversion errors and perform ideal reconstruction. Experimental results show that our approach yields near-perfect reconstructions without sacrificing the editability, thus allowing accurate manipulation of real images. Further, we evaluate the performance of our rectifying network, and see great generalizability towards unseen manipulation types and out-of-domain images.

Peng Zhang, Zhihui Lai, Wenting Chen et al.

Medical vision-language pre-training (VLP) offers significant potential for advancing medical image understanding by leveraging paired image-report data. However, existing methods are limited by Fa}lse Negatives (FaNe) induced by semantically similar texts and insufficient fine-grained cross-modal alignment. To address these limitations, we propose FaNe, a semantic-enhanced VLP framework. To mitigate false negatives, we introduce a semantic-aware positive pair mining strategy based on text-text similarity with adaptive normalization. Furthermore, we design a text-conditioned sparse attention pooling module to enable fine-grained image-text alignment through localized visual representations guided by textual cues. To strengthen intra-modal discrimination, we develop a hard-negative aware contrastive loss that adaptively reweights semantically similar negatives. Extensive experiments on five downstream medical imaging benchmarks demonstrate that FaNe achieves state-of-the-art performance across image classification, object detection, and semantic segmentation, validating the effectiveness of our framework.

Gang Cheng, Xin Gao, Li Hu et al.

We introduce Wan-Animate, a unified framework for character animation and replacement. Given a character image and a reference video, Wan-Animate can animate the character by precisely replicating the expressions and movements of the character in the video to generate high-fidelity character videos. Alternatively, it can integrate the animated character into the reference video to replace the original character, replicating the scene's lighting and color tone to achieve seamless environmental integration. Wan-Animate is built upon the Wan model. To adapt it for character animation tasks, we employ a modified input paradigm to differentiate between reference conditions and regions for generation. This design unifies multiple tasks into a common symbolic representation. We use spatially-aligned skeleton signals to replicate body motion and implicit facial features extracted from source images to reenact expressions, enabling the generation of character videos with high controllability and expressiveness. Furthermore, to enhance environmental integration during character replacement, we develop an auxiliary Relighting LoRA. This module preserves the character's appearance consistency while applying the appropriate environmental lighting and color tone. Experimental results demonstrate that Wan-Animate achieves state-of-the-art performance. We are committed to open-sourcing the model weights and its source code.

Xindian Ma, Wenyuan Liu, Peng Zhang et al.

Inspired by the Bloch Sphere representation, we propose a novel rotary position encoding on a three-dimensional sphere, named 3D Rotary Position Encoding (3D-RPE). 3D-RPE is an advanced version of the widely used 2D Rotary Position Encoding (RoPE), with two major advantages for modeling long contexts: controllable long-term decay and improved position resolution. For controllable long-term decay, 3D-RPE allows for the regulation of long-term decay within the chunk size, ensuring the modeling of relative positional information between tokens at a distant relative position. For enhanced position resolution, 3D-RPE can mitigate the degradation of position resolution caused by position interpolation on RoPE. We have conducted experiments on long-context Natural Language Understanding (NLU) and long-sequence Language Modeling (LM) tasks. From the experimental results, 3D-RPE achieved performance improvements over RoPE, especially in long-context NLU tasks.

Bingchuan Li, Shaofei Cai, Wei Liu et al.

The semantic controllability of StyleGAN is enhanced by unremitting research. Although the existing weak supervision methods work well in manipulating the style codes along one attribute, the accuracy of manipulating multiple attributes is neglected. Multi-attribute representations are prone to entanglement in the StyleGAN latent space, while sequential editing leads to error accumulation. To address these limitations, we design a Dynamic Style Manipulation Network (DyStyle) whose structure and parameters vary by input samples, to perform nonlinear and adaptive manipulation of latent codes for flexible and precise attribute control. In order to efficient and stable optimization of the DyStyle network, we propose a Dynamic Multi-Attribute Contrastive Learning (DmaCL) method: including dynamic multi-attribute contrastor and dynamic multi-attribute contrastive loss, which simultaneously disentangle a variety of attributes from the generative image and latent space of model. As a result, our approach demonstrates fine-grained disentangled edits along multiple numeric and binary attributes. Qualitative and quantitative comparisons with existing style manipulation methods verify the superiority of our method in terms of the multi-attribute control accuracy and identity preservation without compromising photorealism.

Shiming Chen, Peng Zhang, Guo-Sen Xie et al.

Dynamic texture (DT) exhibits statistical stationarity in the spatial domain and stochastic repetitiveness in the temporal dimension, indicating that different frames of DT possess a high similarity correlation that is critical prior knowledge. However, existing methods cannot effectively learn a promising synthesis model for high-dimensional DT from a small number of training data. In this paper, we propose a novel DT synthesis method, which makes full use of similarity prior knowledge to address this issue. Our method bases on the proposed kernel similarity embedding, which not only can mitigate the high-dimensionality and small sample issues, but also has the advantage of modeling nonlinear feature relationship. Specifically, we first raise two hypotheses that are essential for DT model to generate new frames using similarity correlation. Then, we integrate kernel learning and extreme learning machine into a unified synthesis model to learn kernel similarity embedding for representing DT. Extensive experiments on DT videos collected from the internet and two benchmark datasets, i.e., Gatech Graphcut Textures and Dyntex, demonstrate that the learned kernel similarity embedding can effectively exhibit the discriminative representation for DT. Accordingly, our method is capable of preserving the long-term temporal continuity of the synthesized DT sequences with excellent sustainability and generalization. Meanwhile, it effectively generates realistic DT videos with fast speed and low computation, compared with the state-of-the-art methods. The code and more synthesis videos are available at our project page https://shiming-chen.github.io/Similarity-page/Similarit.html.

Peng Zhang, Xinyu Zhu, Zhanzhan Cheng et al.

Fine-grained image recognition has been a hot research topic in computer vision due to its various applications. The-state-of-the-art is the part/region-based approaches that first localize discriminative parts/regions, and then learn their fine-grained features. However, these approaches have some inherent drawbacks: 1) the discriminative feature representation of an object is prone to be disturbed by complicated background; 2) it is unreasonable and inflexible to fix the number of salient parts, because the intended parts may be unavailable under certain circumstances due to occlusion or incompleteness, and 3) the spatial correlation among different salient parts has not been thoroughly exploited (if not completely neglected). To overcome these drawbacks, in this paper we propose a new, simple yet robust method by building part sequence model on the attended object region. Concretely, we first try to alleviate the background effect by using a region attention mechanism to generate the attended region from the original image. Then, instead of localizing different salient parts and extracting their features separately, we learn the part representation implicitly by applying a mapping function on the serialized features of the object. Finally, we combine the region attending network and the part sequence learning network into a unified framework that can be trained end-to-end with only image-level labels. Our extensive experiments on three fine-grained benchmarks show that the proposed method achieves the state of the art performance.

Peng Zhang, Shujian Yu, Jiamiao Xu et al.

It remains a huge challenge to design effective and efficient trackers under complex scenarios, including occlusions, illumination changes and pose variations. To cope with this problem, a promising solution is to integrate the temporal consistency across consecutive frames and multiple feature cues in a unified model. Motivated by this idea, we propose a novel correlation filter-based tracker in this work, in which the temporal relatedness is reconciled under a multi-task learning framework and the multiple feature cues are modeled using a multi-view learning approach. We demonstrate the resulting regression model can be efficiently learned by exploiting the structure of blockwise diagonal matrix. A fast blockwise diagonal matrix inversion algorithm is developed thereafter for efficient online tracking. Meanwhile, we incorporate an adaptive scale estimation mechanism to strengthen the stability of scale variation tracking. We implement our tracker using two types of features and test it on two benchmark datasets. Experimental results demonstrate the superiority of our proposed approach when compared with other state-of-the-art trackers. project homepage http://bmal.hust.edu.cn/project/KMF2JMTtracking.html

Chaojian Yu, Xinyi Zhao, Qi Zheng et al.

Fine-grained visual recognition is challenging because it highly relies on the modeling of various semantic parts and fine-grained feature learning. Bilinear pooling based models have been shown to be effective at fine-grained recognition, while most previous approaches neglect the fact that inter-layer part feature interaction and fine-grained feature learning are mutually correlated and can reinforce each other. In this paper, we present a novel model to address these issues. First, a cross-layer bilinear pooling approach is proposed to capture the inter-layer part feature relations, which results in superior performance compared with other bilinear pooling based approaches. Second, we propose a novel hierarchical bilinear pooling framework to integrate multiple cross-layer bilinear features to enhance their representation capability. Our formulation is intuitive, efficient and achieves state-of-the-art results on the widely used fine-grained recognition datasets.