High perceptual quality and low distortion degree are two important goals in image restoration tasks such as super-resolution (SR). Most of the existing SR methods aim to achieve these goals by minimizing the corresponding yet conflicting losses, such as the $\ell_1$ loss and the adversarial loss. Unfortunately, the commonly used gradient-based optimizers, such as Adam, are hard to balance these objectives due to the opposite gradient decent directions of the contradictory losses. In this paper, we formulate the perception-distortion trade-off in SR as a multi-objective optimization problem and develop a new optimizer by integrating the gradient-free evolutionary algorithm (EA) with gradient-based Adam, where EA and Adam focus on the divergence and convergence of the optimization directions respectively. As a result, a population of optimal models with different perception-distortion preferences is obtained. We then design a fusion network to merge these models into a single stronger one for an effective perception-distortion trade-off. Experiments demonstrate that with the same backbone network, the perception-distortion balanced SR model trained by our method can achieve better perceptual quality than its competitors while attaining better reconstruction fidelity. Codes and models can be found at https://github.com/csslc/EA-Adam.
Aligning large language models(LLMs) with human is a critical step in effectively utilizing their pre-trained capabilities across a wide array of language tasks. Current instruction tuning practices often rely on expanding dataset size without a clear strategy for ensuring data quality, which can inadvertently introduce noise and degrade model performance. To address this challenge, we introduce Nuggets, a novel and efficient methodology that employs one shot learning to select high-quality instruction data from expansive datasets. Nuggets assesses the potential of individual instruction examples to act as effective one shot examples, thereby identifying those that can significantly enhance diverse task performance. Nuggets utilizes a scoring system based on the impact of candidate examples on the perplexity of a diverse anchor set, facilitating the selection of the most beneficial data for instruction tuning. Through rigorous testing on two benchmarks, including MT-Bench and Alpaca-Eval, we demonstrate that instruction tuning with the top 1% of Nuggets-curated examples substantially outperforms conventional methods that use the full dataset. These findings advocate for a data selection paradigm that prioritizes quality, offering a more efficient pathway to align LLMs with humans.
Online video super-resolution (online-VSR) highly relies on an effective alignment module to aggregate temporal information, while the strict latency requirement makes accurate and efficient alignment very challenging. Though much progress has been achieved, most of the existing online-VSR methods estimate the motion fields of each frame separately to perform alignment, which is computationally redundant and ignores the fact that the motion fields of adjacent frames are correlated. In this work, we propose an efficient Temporal Motion Propagation (TMP) method, which leverages the continuity of motion field to achieve fast pixel-level alignment among consecutive frames. Specifically, we first propagate the offsets from previous frames to the current frame, and then refine them in the neighborhood, which significantly reduces the matching space and speeds up the offset estimation process. Furthermore, to enhance the robustness of alignment, we perform spatial-wise weighting on the warped features, where the positions with more precise offsets are assigned higher importance. Experiments on benchmark datasets demonstrate that the proposed TMP method achieves leading online-VSR accuracy as well as inference speed. The source code of TMP can be found at https://github.com/xtudbxk/TMP.
Although there are currently many benchmarks available for evaluating the long context understanding and reasoning capability of large language models, with the expansion of the context window in these models, the existing long context benchmarks are no longer sufficient for evaluating the long context understanding and reasoning capability of large language models. In this paper, we have developed a fresh long context evaluation benchmark, which we name it Marathon in the form of multiple choice questions, inspired by benchmarks such as MMLU, for assessing the long context comprehension capability of large language models quickly, accurately, and objectively. We have evaluated several of the latest and most popular large language models, as well as three recent and effective long context optimization methods, on our benchmark. This showcases the long context reasoning and comprehension capabilities of these large language models and validates the effectiveness of these optimization methods. Marathon is available at https://huggingface.co/datasets/Lemoncoke/Marathon.
Score Distillation Sampling (SDS) has exhibited remarkable performance in conditional 3D content generation. However, a comprehensive understanding of the SDS formulation is still lacking, hindering the development of 3D generation. In this work, we present an interpretation of SDS as a combination of three functional components: mode-disengaging, mode-seeking and variance-reducing terms, and analyze the properties of each. We show that problems such as over-smoothness and color-saturation result from the intrinsic deficiency of the supervision terms and reveal that the variance-reducing term introduced by SDS is sub-optimal. Additionally, we shed light on the adoption of large Classifier-Free Guidance (CFG) scale for 3D generation. Based on the analysis, we propose a simple yet effective approach named Stable Score Distillation (SSD) which strategically orchestrates each term for high-quality 3D generation. Extensive experiments validate the efficacy of our approach, demonstrating its ability to generate high-fidelity 3D content without succumbing to issues such as over-smoothness and over-saturation, even under low CFG conditions with the most challenging NeRF representation.
Building generalist agents that can accomplish many goals in rich open-ended environments is one of the research frontiers for reinforcement learning. A key limiting factor for building generalist agents with RL has been the need for a large number of reward functions for achieving different goals. We investigate the feasibility of using off-the-shelf vision-language models, or VLMs, as sources of rewards for reinforcement learning agents. We show how rewards for visual achievement of a variety of language goals can be derived from the CLIP family of models, and used to train RL agents that can achieve a variety of language goals. We showcase this approach in two distinct visual domains and present a scaling trend showing how larger VLMs lead to more accurate rewards for visual goal achievement, which in turn produces more capable RL agents.
Guided image restoration (GIR), such as guided depth map super-resolution and pan-sharpening, aims to enhance a target image using guidance information from another image of the same scene. Currently, joint image filtering-inspired deep learning-based methods represent the state-of-the-art for GIR tasks. Those methods either deal with GIR in an end-to-end way by elaborately designing filtering-oriented deep neural network (DNN) modules, focusing on the feature-level fusion of inputs; or explicitly making use of the traditional joint filtering mechanism by parameterizing filtering coefficients with DNNs, working on image-level fusion. The former ones are good at recovering contextual information but tend to lose fine-grained details, while the latter ones can better retain textual information but might lead to content distortions. In this work, to inherit the advantages of both methodologies while mitigating their limitations, we proposed a Simultaneous Feature and Image Guided Fusion (SFIGF) network, that simultaneously considers feature and image-level guided fusion following the guided filter (GF) mechanism. In the feature domain, we connect the cross-attention (CA) with GF, and propose a GF-inspired CA module for better feature-level fusion; in the image domain, we fully explore the GF mechanism and design GF-like structure for better image-level fusion. Since guided fusion is implemented in both feature and image domains, the proposed SFIGF is expected to faithfully reconstruct both contextual and textual information from sources and thus lead to better GIR results. We apply SFIGF to 4 typical GIR tasks, and experimental results on these tasks demonstrate its effectiveness and general availability.
This paper presents Audio-Visual LLM, a Multimodal Large Language Model that takes both visual and auditory inputs for holistic video understanding. A key design is the modality-augmented training, which involves the integration of modality-specific tokens engineered to activate the appropriate visual and/or auditory encoder selectively. This mechanism is pivotal in enabling end-to-end joint training with video data at different modalities, including visual-only, audio-only, and audio-visual formats. Moreover, we introduce a high-quality video instruction dataset, derived from GPT-4. This dataset allows Audio-Visual LLM to adeptly process a variety of task-oriented video instructions, ranging from multi-turn conversations and audio-visual narratives to complex reasoning tasks. Extensive experiments demonstrate that Audio-Visual LLM impressively achieves strong zero-shot results across a range of video understanding tasks. For example, Audio-Visual LLM achieves an accuracy of 53.7% on MSRVTT-QA, outperforming non-LLM-based InterVideo by 6.6% and LLM-based Valley by 4.4%, respectively. Additionally, our Audio-Visual LLM also achieves competitive performance on audio tasks (e.g., AudioCaps).
The massive growth of image-text data through web crawling inherently presents the challenge of variability in data quality. This paper introduces a novel algorithm, rooted in human knowledge, to compress this vast corpus of web-crawled image-text datasets to a compact and high-quality form. Our method unfolds in three major steps. First, we collect an image-text dataset, wherein each image is associated with multiple captions sourced from diverse origins. Then, to systemically capture human preferences regarding the best caption paired with each image, we establish a comprehensive set of both subjective and objective criteria for critically guiding the alignment assessment from labelers. Lastly, we train a reward model on the annotated dataset to internalize the nuanced human understanding of image-text alignment. The resulting reward model thus can act as a human-like referee to filter misaligned/low-quality image-text pairs. Extensive experiments demonstrate that we are able to secure (or even improve) model performance by compressing the image-text datasets up to ~90%. An impressive example is that, by aggressively reducing the total training sample from 130M to 15.5M (e.g., ~9x smaller), our BLIP-B/16 models still consistently show superior performance compared with the full-size-dataset counterpart on image-text retrieval (Flickr30K, COCO) by ~2.5% in Recall@1, and on image-captioning (Nocaps, COCO) by ~10.0% in CIDEr and ~2.7% in SPICE.
In this paper, we introduce a novel approach to novel object captioning which employs relative contrastive learning to learn visual and semantic alignment. Our approach maximizes compatibility between regions and object tags in a contrastive manner. To set up a proper contrastive learning objective, for each image, we augment tags by leveraging the relative nature of positive and negative pairs obtained from foundation models such as CLIP. We then use the rank of each augmented tag in a list as a relative relevance label to contrast each top-ranked tag with a set of lower-ranked tags. This learning objective encourages the top-ranked tags to be more compatible with their image and text context than lower-ranked tags, thus improving the discriminative ability of the learned multi-modality representation. We evaluate our approach on two datasets and show that our proposed RCA-NOC approach outperforms state-of-the-art methods by a large margin, demonstrating its effectiveness in improving vision-language representation for novel object captioning.