Hallucinations and unfaithful synthesis due to inaccurate prompts with insufficient semantic details are widely observed in multimodal generative models. A prevalent strategy to align multiple modalities is to fine-tune the generator with a large number of annotated text-image pairs. However, such a procedure is labor-consuming and resource-draining. The key question we ask is: can we enhance the quality and faithfulness of text-driven generative models beyond extensive text-image pair annotations? To address this question, we propose Knowledge Pursuit Prompting (KPP), a zero-shot framework that iteratively incorporates external knowledge to help generators produce reliable visual content. Instead of training generators to handle generic prompts, KPP employs a recursive knowledge query process to gather informative external facts from the knowledge base, instructs a language model to compress the acquired knowledge for prompt refinement, and utilizes text-driven generators for visual synthesis. The entire process is zero-shot, without accessing the architectures and parameters of generative models. We evaluate the framework across multiple text-driven generative tasks (image, 3D rendering, and video) on datasets of different domains. We further demonstrate the extensibility and adaptability of KPP through varying foundation model bases and instructions. Our results show that KPP is capable of generating faithful and semantically rich content across diverse visual domains, offering a promising solution to improve multimodal generative models.
We introduce PoseGPT, a framework employing Large Language Models (LLMs) to understand and reason about 3D human poses from images or textual descriptions. Our work is motivated by the human ability to intuitively understand postures from a single image or a brief description, a process that intertwines image interpretation, world knowledge, and an understanding of body language. Traditional human pose estimation methods, whether image-based or text-based, often lack holistic scene comprehension and nuanced reasoning, leading to a disconnect between visual data and its real-world implications. PoseGPT addresses these limitations by embedding SMPL poses as a distinct signal token within a multi-modal LLM, enabling direct generation of 3D body poses from both textual and visual inputs. This approach not only simplifies pose prediction but also empowers LLMs to apply their world knowledge in reasoning about human poses, fostering two advanced tasks: speculative pose generation and reasoning about pose estimation. These tasks involve reasoning about humans to generate 3D poses from subtle text queries, possibly accompanied by images. We establish benchmarks for these tasks, moving beyond traditional 3D pose generation and estimation methods. Our results show that PoseGPT outperforms existing multimodal LLMs and task-sepcific methods on these newly proposed tasks. Furthermore, PoseGPT's ability to understand and generate 3D human poses based on complex reasoning opens new directions in human pose analysis.
Graph neural networks (GNNs) present a promising alternative to CNNs and transformers in certain image processing applications due to their parameter-efficiency in modeling spatial relationships. Currently, a major area of research involves the converting non-graph input data for GNN-based models, notably in scenarios where the data originates from images. One approach involves converting images into nodes by identifying significant keypoints within them. Super-Retina, a semi-supervised technique, has been utilized for detecting keypoints in retinal images. However, its limitations lie in the dependency on a small initial set of ground truth keypoints, which is progressively expanded to detect more keypoints. Having encountered difficulties in detecting consistent initial keypoints in brain images using SIFT and LoFTR, we proposed a new approach: radiomic feature-based keypoint detection. Demonstrating the anatomical significance of the detected keypoints was achieved by showcasing their efficacy in improving registration processes guided by these keypoints. Subsequently, these keypoints were employed as the ground truth for the keypoint detection method (LK-SuperRetina). Furthermore, the study showcases the application of GNNs in image matching, highlighting their superior performance in terms of both the number of good matches and confidence scores. This research sets the stage for expanding GNN applications into various other applications, including but not limited to image classification, segmentation, and registration.
Chinese calligraphy can be viewed as a unique form of visual art. Recent advancements in computer vision hold significant potential for the future development of generative models in the realm of Chinese calligraphy. Nevertheless, methods of Chinese calligraphy inpainting, which can be effectively used in the art and education fields, remain relatively unexplored. In this paper, we introduce a new model that harnesses recent advancements in both Chinese calligraphy generation and image inpainting. We demonstrate that our proposed model CalliPaint can produce convincing Chinese calligraphy.
Image captioning studies heavily rely on automatic evaluation metrics such as BLEU and METEOR. However, such n-gram-based metrics have been shown to correlate poorly with human evaluation, leading to the proposal of alternative metrics such as SPICE for English; however, no equivalent metrics have been established for other languages. Therefore, in this study, we propose an automatic evaluation metric called JaSPICE, which evaluates Japanese captions based on scene graphs. The proposed method generates a scene graph from dependencies and the predicate-argument structure, and extends the graph using synonyms. We conducted experiments employing 10 image captioning models trained on STAIR Captions and PFN-PIC and constructed the Shichimi dataset, which contains 103,170 human evaluations. The results showed that our metric outperformed the baseline metrics for the correlation coefficient with the human evaluation.
It is well-known that training neural networks for image classification with empirical risk minimization (ERM) makes them vulnerable to relying on spurious attributes instead of causal ones for prediction. Previously, deep feature re-weighting (DFR) has proposed retraining the last layer of a pre-trained network on balanced data concerning spurious attributes, making it robust to spurious correlation. However, spurious attribute annotations are not always available. In order to provide group robustness without such annotations, we propose a new method, called loss-based feature re-weighting (LFR), in which we infer a grouping of the data by evaluating an ERM-pre-trained model on a small left-out split of the training data. Then, a balanced number of samples is chosen by selecting high-loss samples from misclassified data points and low-loss samples from correctly-classified ones. Finally, we retrain the last layer on the selected balanced groups to make the model robust to spurious correlation. For a complete assessment, we evaluate LFR on various versions of Waterbirds and CelebA datasets with different spurious correlations, which is a novel technique for observing the model's performance in a wide range of spuriosity rates. While LFR is extremely fast and straightforward, it outperforms the previous methods that do not assume group label availability, as well as the DFR with group annotations provided, in cases of high spurious correlation in the training data.
Fast and accurate segmentation of neonatal brain images is highly desired to better understand and detect changes during development and disease. Yet, the limited availability of ground truth datasets, lack of standardized acquisition protocols, and wide variations of head positioning pose challenges for method development. A few automated image analysis pipelines exist for newborn brain MRI segmentation, but they often rely on time-consuming procedures and require resampling to a common resolution, subject to loss of information due to interpolation and down-sampling. Without registration and image resampling, variations with respect to head positions and voxel resolutions have to be addressed differently. In deep-learning, external augmentations are traditionally used to artificially expand the representation of spatial variability, increasing the training dataset size and robustness. However, these transformations in the image space still require resampling, reducing accuracy specifically in the context of label interpolation. We recently introduced the concept of resolution-independence with the Voxel-size Independent Neural Network framework, VINN. Here, we extend this concept by additionally shifting all rigid-transforms into the network architecture with a four degree of freedom (4-DOF) transform module, enabling resolution-aware internal augmentations (VINNA). In this work we show that VINNA (i) significantly outperforms state-of-the-art external augmentation approaches, (ii) effectively addresses the head variations present specifically in newborn datasets, and (iii) retains high segmentation accuracy across a range of resolutions (0.5-1.0 mm). The 4-DOF transform module is a powerful, general approach to implement spatial augmentation without requiring image or label interpolation. The specific network application to newborns will be made publicly available as VINNA4neonates.
We introduce HiDiffusion, a tuning-free framework comprised of Resolution-Aware U-Net (RAU-Net) and Modified Shifted Window Multi-head Self-Attention (MSW-MSA) to enable pretrained large text-to-image diffusion models to efficiently generate high-resolution images (e.g. 1024$\times$1024) that surpass the training image resolution. Pretrained diffusion models encounter unreasonable object duplication in generating images beyond the training image resolution. We attribute it to the mismatch between the feature map size of high-resolution images and the receptive field of U-Net's convolution. To address this issue, we propose a simple yet scalable method named RAU-Net. RAU-Net dynamically adjusts the feature map size to match the convolution's receptive field in the deep block of U-Net. Another obstacle in high-resolution synthesis is the slow inference speed of U-Net. Our observations reveal that the global self-attention in the top block, which exhibits locality, however, consumes the majority of computational resources. To tackle this issue, we propose MSW-MSA. Unlike previous window attention mechanisms, our method uses a much larger window size and dynamically shifts windows to better accommodate diffusion models. Extensive experiments demonstrate that our HiDiffusion can scale diffusion models to generate 1024$\times$1024, 2048$\times$2048, or even 4096$\times$4096 resolution images, while simultaneously reducing inference time by 40\%-60\%, achieving state-of-the-art performance on high-resolution image synthesis. The most significant revelation of our work is that a pretrained diffusion model on low-resolution images is scalable for high-resolution generation without further tuning. We hope this revelation can provide insights for future research on the scalability of diffusion models.
We introduce multi-slice reasoning, a new notion for single-view 3D reconstruction which challenges the current and prevailing belief that multi-view synthesis is the most natural conduit between single-view and 3D. Our key observation is that object slicing is more advantageous than altering views to reveal occluded structures. Specifically, slicing is more occlusion-revealing since it can peel through any occluders without obstruction. In the limit, i.e., with infinitely many slices, it is guaranteed to unveil all hidden object parts. We realize our idea by developing Slice3D, a novel method for single-view 3D reconstruction which first predicts multi-slice images from a single RGB image and then integrates the slices into a 3D model using a coordinate-based transformer network for signed distance prediction. The slice images can be regressed or generated, both through a U-Net based network. For the former, we inject a learnable slice indicator code to designate each decoded image into a spatial slice location, while the slice generator is a denoising diffusion model operating on the entirety of slice images stacked on the input channels. We conduct extensive evaluation against state-of-the-art alternatives to demonstrate superiority of our method, especially in recovering complex and severely occluded shape structures, amid ambiguities. All Slice3D results were produced by networks trained on a single Nvidia A40 GPU, with an inference time less than 20 seconds.
Diffusion models have achieved remarkable success in generating realistic images but suffer from generating accurate human hands, such as incorrect finger counts or irregular shapes. This difficulty arises from the complex task of learning the physical structure and pose of hands from training images, which involves extensive deformations and occlusions. For correct hand generation, our paper introduces a lightweight post-processing solution called $\textbf{HandRefiner}$. HandRefiner employs a conditional inpainting approach to rectify malformed hands while leaving other parts of the image untouched. We leverage the hand mesh reconstruction model that consistently adheres to the correct number of fingers and hand shape, while also being capable of fitting the desired hand pose in the generated image. Given a generated failed image due to malformed hands, we utilize ControlNet modules to re-inject such correct hand information. Additionally, we uncover a phase transition phenomenon within ControlNet as we vary the control strength. It enables us to take advantage of more readily available synthetic data without suffering from the domain gap between realistic and synthetic hands. Experiments demonstrate that HandRefiner can significantly improve the generation quality quantitatively and qualitatively. The code is available at https://github.com/wenquanlu/HandRefiner .