用GPT-4训练微型机器人游泳；个性化图像生成；动态纹理迁移；光场合成

本文首发于公众号：机器感知

用GPT-4训练微型机器人游泳；个性化图像生成；动态纹理迁移；光场合成

Training microrobots to swim by a large language model

Machine learning and artificial intelligence have recently represented a popular paradigm for designing and optimizing robotic systems across various scales. Recent studies have showcased the innovative application of large language models (LLMs) in industrial control [1] and in directing legged walking robots [2]. In this study, we utilize an LLM, GPT-4, to train two prototypical microrobots for swimming in viscous fluids. Adopting a few-shot learning approach, we develop a minimal, unified prompt composed of only five sentences. The same concise prompt successfully guides two distinct articulated microrobots -- the three-link swimmer and the three-sphere swimmer -- in mastering their signature strokes. These strokes, initially conceptualized by physicists, are now effectively interpreted and applied by the LLM, enabling the microrobots to circumvent the physical constraints inherent to micro-locomotion. Remarkably, our LLM-based decision-making strategy substantially surpasses a traditional reinforcement learning method in terms of training speed.

Merging Multi-Task Models via Weight-Ensembling Mixture of Experts

Merging various task-specific Transformer-based models trained on different tasks into a single unified model can execute all the tasks concurrently. Previous methods, exemplified by task arithmetic, have been proven to be both effective and scalable. Existing methods have primarily focused on seeking a static optimal solution within the original model parameter space. A notable challenge is mitigating the interference between parameters of different models, which can substantially deteriorate performance. In this paper, we propose to merge most of the parameters while upscaling the MLP of the Transformer layers to a weight-ensembling mixture of experts (MoE) module, which can dynamically integrate shared and task-specific knowledge based on the input, thereby providing a more flexible solution that can adapt to the specific needs of each instance. Our key insight is that by identifying and separating shared knowledge and task-specific knowledge, and then dynamically integrating them, we can mitigate the parameter interference problem to a great extent.

EE-Tuning: An Economical yet Scalable Solution for Tuning Early-Exit Large Language Models

This work introduces EE-Tuning, a lightweight and economical solution to training/tuning early-exit large language models (LLMs). In contrast to the common approach of full-parameter pre-training, EE-Tuning augments any pre-trained (and possibly fine-tuned) standard LLM with additional early-exit layers that are tuned in a parameter-efficient manner, which requires significantly less computational resources and training data. Our implementation of EE-Tuning achieves outstanding training efficiency via extensive performance optimizations, as well as scalability due to its full compatibility with 3D parallelism. Results of systematic experiments validate the efficacy of EE-Tuning, confirming that effective early-exit LLM inference can be achieved with a limited training budget.

Diffusion-based Light Field Synthesis

Light fields (LFs), conducive to comprehensive scene radiance recorded across angular dimensions, find wide applications in 3D reconstruction, virtual reality, and computational photography.However, the LF acquisition is inevitably time-consuming and resource-intensive due to the mainstream acquisition strategy involving manual capture or laborious software synthesis.Given such a challenge, we introduce LFdiff, a straightforward yet effective diffusion-based generative framework tailored for LF synthesis, which adopts only a single RGB image as input.LFdiff leverages disparity estimated by a monocular depth estimation network and incorporates two distinctive components: a novel condition scheme and a noise estimation network tailored for LF data.Specifically, we design a position-aware warping condition scheme, enhancing inter-view geometry learning via a robust conditional signal. We then propose DistgUnet, a disentanglement-based noise estimation network, to harness comprehensive LF representations.

Dynamic Texture Transfer using PatchMatch and Transformers

How to automatically transfer the dynamic texture of a given video to the target still image is a challenging and ongoing problem. In this paper, we propose to handle this task via a simple yet effective model that utilizes both PatchMatch and Transformers. The key idea is to decompose the task of dynamic texture transfer into two stages, where the start frame of the target video with the desired dynamic texture is synthesized in the first stage via a distance map guided texture transfer module based on the PatchMatch algorithm. Then, in the second stage, the synthesized image is decomposed into structure-agnostic patches, according to which their corresponding subsequent patches can be predicted by exploiting the powerful capability of Transformers equipped with VQ-VAE for processing long discrete sequences. After getting all those patches, we apply a Gaussian weighted average merging strategy to smoothly assemble them into each frame of the target stylized video.

SeFi-IDE: Semantic-Fidelity Identity Embedding for Personalized Diffusion-Based Generation

Advanced diffusion-based Text-to-Image (T2I) models, such as the Stable Diffusion Model, have made significant progress in generating diverse and high-quality images using text prompts alone. However, T2I models are unable to accurately map identities (IDs) when non-famous users require personalized image generation. The main problem is that existing T2I models do not learn the ID-image alignments of new users. The previous methods either failed to accurately fit the face region or lost the interactive generative ability with other existing concepts in T2I models (i.e., unable to generate other concepts described in given prompts such as scenes, actions, and facial attributes). In this paper, we focus on accurate and semantic-fidelity ID embedding into the Stable Diffusion Model for personalized generation. We address this challenge from two perspectives: face-wise region fitting, and semantic-fidelity token optimization. Specifically, we first visualize the attention overfit problem, and propose a face-wise attention loss to fit the face region instead of the whole target image. This key trick significantly enhances the ID accuracy and interactive generative ability with other existing concepts. Then, we optimize one ID representation as multiple per-stage tokens where each token contains two disentangled features. This expansion of the textual conditioning space enhances semantic-fidelity control.