Chen Li

Chen Li, Jiahao Lin, Gim Hee Lee

In this paper, we tackle the challenging task of learning a generalizable human NeRF model from a monocular video. Although existing generalizable human NeRFs have achieved impressive results, they require muti-view images or videos which might not be always available. On the other hand, some works on free-viewpoint rendering of human from monocular videos cannot be generalized to unseen identities. In view of these limitations, we propose GHuNeRF to learn a generalizable human NeRF model from a monocular video of the human performer. We first introduce a visibility-aware aggregation scheme to compute vertex-wise features, which is used to construct a 3D feature volume. The feature volume can only represent the overall geometry of the human performer with insufficient accuracy due to the limited resolution. To solve this, we further enhance the volume feature with temporally aligned point-wise features using an attention mechanism. Finally, the enhanced feature is used for predicting density and color for each sampled point. A surface-guided sampling strategy is also adopted to improve the efficiency for both training and inference. We validate our approach on the widely-used ZJU-MoCap dataset, where we achieve comparable performance with existing multi-view video based approaches. We also test on the monocular People-Snapshot dataset and achieve better performance than existing works when only monocular video is used. Our code is available at the project website.

Pengze Zhang, Hubery Yin, Chen Li et al.

Continuous diffusion models are commonly acknowledged to display a deterministic probability flow, whereas discrete diffusion models do not. In this paper, we aim to establish the fundamental theory for the probability flow of discrete diffusion models. Specifically, we first prove that the continuous probability flow is the Monge optimal transport map under certain conditions, and also present an equivalent evidence for discrete cases. In view of these findings, we are then able to define the discrete probability flow in line with the principles of optimal transport. Finally, drawing upon our newly established definitions, we propose a novel sampling method that surpasses previous discrete diffusion models in its ability to generate more certain outcomes. Extensive experiments on the synthetic toy dataset and the CIFAR-10 dataset have validated the effectiveness of our proposed discrete probability flow. Code is released at: https://github.com/PangzeCheung/Discrete-Probability-Flow.

Yuying Ge, Sijie Zhao, Chen Li et al. · tencent-ai

In this technical report, we introduce SEED-Data-Edit: a unique hybrid dataset for instruction-guided image editing, which aims to facilitate image manipulation using open-form language. SEED-Data-Edit is composed of three distinct types of data: (1) High-quality editing data produced by an automated pipeline, ensuring a substantial volume of diverse image editing pairs. (2) Real-world scenario data collected from the internet, which captures the intricacies of user intentions for promoting the practical application of image editing in the real world. (3) High-precision multi-turn editing data annotated by humans, which involves multiple rounds of edits for simulating iterative editing processes. The combination of these diverse data sources makes SEED-Data-Edit a comprehensive and versatile dataset for training language-guided image editing model. We fine-tune a pretrained Multimodal Large Language Model (MLLM) that unifies comprehension and generation with SEED-Data-Edit. The instruction tuned model demonstrates promising results, indicating the potential and effectiveness of SEED-Data-Edit in advancing the field of instructional image editing. The datasets are released in https://huggingface.co/datasets/AILab-CVC/SEED-Data-Edit.

Buzhen Huang, Chen Li, Chongyang Xu et al.

Existing multi-person human reconstruction approaches mainly focus on recovering accurate poses or avoiding penetration, but overlook the modeling of close interactions. In this work, we tackle the task of reconstructing closely interactive humans from a monocular video. The main challenge of this task comes from insufficient visual information caused by depth ambiguity and severe inter-person occlusion. In view of this, we propose to leverage knowledge from proxemic behavior and physics to compensate the lack of visual information. This is based on the observation that human interaction has specific patterns following the social proxemics. Specifically, we first design a latent representation based on Vector Quantised-Variational AutoEncoder (VQ-VAE) to model human interaction. A proxemics and physics guided diffusion model is then introduced to denoise the initial distribution. We design the diffusion model as dual branch with each branch representing one individual such that the interaction can be modeled via cross attention. With the learned priors of VQ-VAE and physical constraint as the additional information, our proposed approach is capable of estimating accurate poses that are also proxemics and physics plausible. Experimental results on Hi4D, 3DPW, and CHI3D demonstrate that our method outperforms existing approaches. The code is available at \url{https://github.com/boycehbz/HumanInteraction}.

Chen Li, Bipin Rajendran

Recent strides in low-latency spiking neural network (SNN) algorithms have drawn significant interest, particularly due to their event-driven computing nature and fast inference capability. One of the most efficient ways to construct a low-latency SNN is by converting a pre-trained, low-bit artificial neural network (ANN) into an SNN. However, this conversion process faces two main challenges: First, converting SNNs from low-bit ANNs can lead to ``occasional noise" -- the phenomenon where occasional spikes are generated in spiking neurons where they should not be -- during inference, which significantly lowers SNN accuracy. Second, although low-latency SNNs initially show fast improvements in accuracy with time steps, these accuracy growths soon plateau, resulting in their peak accuracy lagging behind both full-precision ANNs and traditional ``long-latency SNNs'' that prioritize precision over speed. In response to these two challenges, this paper introduces a novel technique named ``noise adaptor.'' Noise adaptor can model occasional noise during training and implicitly optimize SNN accuracy, particularly at high simulation times $T$. Our research utilizes the ResNet model for a comprehensive analysis of the impact of the noise adaptor on low-latency SNNs. The results demonstrate that our method outperforms the previously reported quant-ANN-to-SNN conversion technique. We achieved an accuracy of 95.95\% within 4 time steps on CIFAR-10 using ResNet-18, and an accuracy of 74.37\% within 64 time steps on ImageNet using ResNet-50. Remarkably, these results were obtained without resorting to any noise correction methods during SNN inference, such as negative spikes or two-stage SNN simulations. Our approach significantly boosts the peak accuracy of low-latency SNNs, bringing them on par with the accuracy of full-precision ANNs. Code will be open source.

Pengze Zhang, Hubery Yin, Chen Li et al.

Most diffusion models assume that the reverse process adheres to a Gaussian distribution. However, this approximation has not been rigorously validated, especially at singularities, where t=0 and t=1. Improperly dealing with such singularities leads to an average brightness issue in applications, and limits the generation of images with extreme brightness or darkness. We primarily focus on tackling singularities from both theoretical and practical perspectives. Initially, we establish the error bounds for the reverse process approximation, and showcase its Gaussian characteristics at singularity time steps. Based on this theoretical insight, we confirm the singularity at t=1 is conditionally removable while it at t=0 is an inherent property. Upon these significant conclusions, we propose a novel plug-and-play method SingDiffusion to address the initial singular time step sampling, which not only effectively resolves the average brightness issue for a wide range of diffusion models without extra training efforts, but also enhances their generation capability in achieving notable lower FID scores.

Run Yu, Yangdi Liu, Wen-Da Wei et al.

Recently,vision-based robotic manipulation has garnered significant attention and witnessed substantial advancements. 2D image-based and 3D point cloud-based policy learning represent two predominant paradigms in the field, with recent studies showing that the latter consistently outperforms the former in terms of both policy performance and generalization, thereby underscoring the value and significance of 3D information. However, 3D point cloud-based approaches face the significant challenge of high data acquisition costs, limiting their scalability and real-world deployment. To address this issue, we propose a novel framework NoReal3D: which introduces the 3DStructureFormer, a learnable 3D perception module capable of transforming monocular images into geometrically meaningful pseudo-point cloud features, effectively fused with the 2D encoder output features. Specially, the generated pseudo-point clouds retain geometric and topological structures so we design a pseudo-point cloud encoder to preserve these properties, making it well-suited for our framework. We also investigate the effectiveness of different feature fusion strategies.Our framework enhances the robot's understanding of 3D spatial structures while completely eliminating the substantial costs associated with 3D point cloud acquisition.Extensive experiments across various tasks validate that our framework can achieve performance comparable to 3D point cloud-based methods, without the actual point cloud data.

Chen Li, Elena Ferro, Corey Lammie et al.

Analog In-Memory Computing (AIMC) offers a promising solution to the von Neumann bottleneck. However, deploying transformer models on AIMC remains challenging due to their inherent need for flexibility and adaptability across diverse tasks. For the benefits of AIMC to be fully realized, weights of static vector-matrix multiplications must be mapped and programmed to analog devices in a weight-stationary manner. This poses two challenges for adapting a base network to hardware and downstream tasks: (i) conventional analog hardware-aware (AHWA) training requires retraining the entire model, and (ii) reprogramming analog devices is both time- and energy-intensive. To address these issues, we propose Analog Hardware-Aware Low-Rank Adaptation (AHWA-LoRA) training, a novel approach for efficiently adapting transformers to AIMC hardware. AHWA-LoRA training keeps the analog weights fixed as meta-weights and introduces lightweight external LoRA modules for both hardware and task adaptation. We validate AHWA-LoRA training on SQuAD v1.1 and the GLUE benchmark, demonstrate its scalability to larger models, and show its effectiveness in instruction tuning and reinforcement learning. We further evaluate a practical deployment scenario that balances AIMC tile latency with digital LoRA processing using optimized pipeline strategies, with RISC-V-based programmable multi-core accelerators. This hybrid architecture achieves efficient transformer inference with only a 4% per-layer overhead compared to a fully AIMC implementation.