Ziyue Zhang

Bowen Duan, Cong Guo, Chiyue Wei et al.

Large Language Models (LLMs) have achieved impressive performance across diverse domains but remain inefficient during the autoregressive decoding phase. Unlike the prefill stage, which employs compute-bound GEMM operations, decoding executes a sequence of small GEMV-like computations that are memory-bound and underutilize modern accelerators. Weight-only vector quantization (VQ) has emerged as an effective compression technique that clusters model weights into a shared codebook and replaces the original weight matrix with low-precision indices, enabling 2-bit-level weight compression. While this approach substantially reduces model size and memory bandwidth, it still suffers from two critical inefficiencies: the low utilization of GEMV computation and frequent memory conflicts during codebook lookups. This paper presents EVA, an efficient vector-quantization-based architecture that addresses both computational and memory bottlenecks in LLM decoding. EVA builds on a simple yet effective insight that combines input-codebook computation with conflict-free memory access. Instead of reconstructing quantized weights from indices, EVA directly performs dot products between input vectors and the weight codebook, transforming LLM decoding from GEMV to GEMM computation. It then performs structured lookups from an intermediate output buffer, eliminating memory bank conflicts. We further design a hardware-software co-optimized architecture specialized for LLM decoding while remaining compatible with conventional prefill execution. Evaluations show that EVA achieves up to 11.17$\times$ speedup and 7.17$\times$ higher energy efficiency compared with the SOTA lookup-based architecture, while preserving arithmetic precision after vector quantization. Our code is available at https://github.com/dbw6/Eva.git.

Xuan Sun, Ziyue Zhang, Guannan Chen et al.

In recent years deep learning methods have shown great superiority in compressed video quality enhancement tasks. Existing methods generally take the raw video as the ground truth and extract practical information from consecutive frames containing various artifacts. However, they do not fully exploit the valid information of compressed and raw videos to guide the quality enhancement for compressed videos. In this paper, we propose a unique Valid Information Guidance scheme (VIG) to enhance the quality of compressed videos by mining valid information from both compressed videos and raw videos. Specifically, we propose an efficient framework, Compressed Redundancy Filtering (CRF) network, to balance speed and enhancement. After removing the redundancy by filtering the information, CRF can use the valid information of the compressed video to reconstruct the texture. Furthermore, we propose a progressive Truth Guidance Distillation (TGD) strategy, which does not need to design additional teacher models and distillation loss functions. By only using the ground truth as input to guide the model to aggregate the correct spatio-temporal correspondence across the raw frames, TGD can significantly improve the enhancement effect without increasing the extra training cost. Extensive experiments show that our method achieves the state-of-the-art performance of compressed video quality enhancement in terms of accuracy and efficiency.

Ziyue Zhang, Luxi Lin, Xiaolin Hu et al.

Diffusion inversion is a task of recovering the noise of an image in a diffusion model, which is vital for controllable diffusion image editing. At present, diffusion inversion still remains a challenging task due to the lack of viable supervision signals. Thus, most existing methods resort to approximation-based solutions, which however are often at the cost of performance or efficiency. To remedy these shortcomings, we propose a novel self-supervised diffusion inversion approach in this paper, termed Deep Inversion (DeepInv). Instead of requiring ground-truth noise annotations, we introduce a self-supervised objective as well as a data augmentation strategy to generate high-quality pseudo noises from real images without manual intervention. Based on these two innovative designs, DeepInv is also equipped with an iterative and multi-scale training regime to train a parameterized inversion solver, thereby achieving the fast and accurate image-to-noise mapping. To the best of our knowledge, this is the first attempt of presenting a trainable solver to predict inversion noise step by step. The extensive experiments show that our DeepInv can achieve much better performance and inference speed than the compared methods, e.g., +40.435% SSIM than EasyInv and +9887.5% speed than ReNoise on COCO dataset. Moreover, our careful designs of trainable solvers can also provide insights to the community. Codes and model parameters will be released in https://github.com/potato-kitty/DeepInv.

Ziyue Zhang, Yuxin Zhang, Quanjian Song et al.

We introduce ObjectAdd, a training-free diffusion modification method to add user-expected objects into user-specified area. The motive of ObjectAdd stems from: first, describing everything in one prompt can be difficult, and second, users often need to add objects into the generated image. To accommodate with real world, our ObjectAdd maintains accurate image consistency after adding objects with technical innovations in: (1) embedding-level concatenation to ensure correct text embedding coalesce; (2) object-driven layout control with latent and attention injection to ensure objects accessing user-specified area; (3) prompted image inpainting in an attention refocusing & object expansion fashion to ensure rest of the image stays the same. With a text-prompted image, our ObjectAdd allows users to specify a box and an object, and achieves: (1) adding object inside the box area; (2) exact content outside the box area; (3) flawless fusion between the two areas

Ziyue Zhang, Shuai Jiang, Congzhentao Huang et al.

RGB-Infrared (RGB-IR) person re-identification (ReID) is a technology where the system can automatically identify the same person appearing at different parts of a video when light is unavailable. The critical challenge of this task is the cross-modality gap of features under different modalities. To solve this challenge, we proposed a Teacher-Student GAN model (TS-GAN) to adopt different domains and guide the ReID backbone to learn better ReID information. (1) In order to get corresponding RGB-IR image pairs, the RGB-IR Generative Adversarial Network (GAN) was used to generate IR images. (2) To kick-start the training of identities, a ReID Teacher module was trained under IR modality person images, which is then used to guide its Student counterpart in training. (3) Likewise, to better adapt different domain features and enhance model ReID performance, three Teacher-Student loss functions were used. Unlike other GAN based models, the proposed model only needs the backbone module at the test stage, making it more efficient and resource-saving. To showcase our model's capability, we did extensive experiments on the newly-released SYSU-MM01 RGB-IR Re-ID benchmark and achieved superior performance to the state-of-the-art with 49.8% Rank-1 and 47.4% mAP.

Ziyue Zhang, Richard YD Xu, Shuai Jiang et al.

Most existing works in Person Re-identification (ReID) focus on settings where illumination either is kept the same or has very little fluctuation. However, the changes in the illumination degree may affect the robustness of a ReID algorithm significantly. To address this problem, we proposed a Two-Stream Network that can separate ReID features from lighting features to enhance ReID performance. Its innovations are threefold: (1) A discriminative entropy loss to ensure the ReID features contain no lighting information. (2) A ReID Teacher model trained by images under "neutral" lighting conditions to guide ReID classification. (3) An illumination Teacher model trained by the differences between the illumination-adjusted and original images to guide illumination classification. We construct two augmented datasets by synthetically changing a set of predefined lighting conditions in two of the most popular ReID benchmarks: Market1501 and DukeMTMC-ReID. Experiments demonstrate that our algorithm outperforms other state-of-the-art works and particularly potent in handling images under extremely low light.