Text-guided diffusion models (TDMs) are widely applied but can fail unexpectedly. Common failures include: (i) natural-looking text prompts generating images with the wrong content, or (ii) different random samples of the latent variables that generate vastly different, and even unrelated, outputs despite being conditioned on the same text prompt. In this work, we aim to study and understand the failure modes of TDMs in more detail. To achieve this, we propose SAGE, an adversarial attack on TDMs that uses image classifiers as surrogate loss functions, to search over the discrete prompt space and the high-dimensional latent space of TDMs to automatically discover unexpected behaviors and failure cases in the image generation. We make several technical contributions to ensure that SAGE finds failure cases of the diffusion model, rather than the classifier, and verify this in a human study. Our study reveals four intriguing properties of TDMs that have not been systematically studied before: (1) We find a variety of natural text prompts producing images that fail to capture the semantics of input texts. We categorize these failures into ten distinct types based on the underlying causes. (2) We find samples in the latent space (which are not outliers) that lead to distorted images independent of the text prompt, suggesting that parts of the latent space are not well-structured. (3) We also find latent samples that lead to natural-looking images which are unrelated to the text prompt, implying a potential misalignment between the latent and prompt spaces. (4) By appending a single adversarial token embedding to an input prompt we can generate a variety of specified target objects, while only minimally affecting the CLIP score. This demonstrates the fragility of language representations and raises potential safety concerns.
Deep supervision, which involves extra supervisions to the intermediate features of a neural network, was widely used in image classification in the early deep learning era since it significantly reduces the training difficulty and eases the optimization like avoiding gradient vanish over the vanilla training. Nevertheless, with the emergence of normalization techniques and residual connection, deep supervision in image classification was gradually phased out. In this paper, we revisit deep supervision for masked image modeling (MIM) that pre-trains a Vision Transformer (ViT) via a mask-and-predict scheme. Experimentally, we find that deep supervision drives the shallower layers to learn more meaningful representations, accelerates model convergence, and expands attention diversities. Our approach, called DeepMIM, significantly boosts the representation capability of each layer. In addition, DeepMIM is compatible with many MIM models across a range of reconstruction targets. For instance, using ViT-B, DeepMIM on MAE achieves 84.2 top-1 accuracy on ImageNet, outperforming MAE by +0.6. By combining DeepMIM with a stronger tokenizer CLIP, our model achieves state-of-the-art performance on various downstream tasks, including image classification (85.6 top-1 accuracy on ImageNet-1K, outperforming MAE-CLIP by +0.8), object detection (52.8 APbox on COCO) and semantic segmentation (53.1 mIoU on ADE20K). Code and models are available at https://github.com/OliverRensu/DeepMIM.
How to generate the ground-truth (GT) image is a critical issue for training realistic image super-resolution (Real-ISR) models. Existing methods mostly take a set of high-resolution (HR) images as GTs and apply various degradations to simulate their low-resolution (LR) counterparts. Though great progress has been achieved, such an LR-HR pair generation scheme has several limitations. First, the perceptual quality of HR images may not be high enough, limiting the quality of Real-ISR outputs. Second, existing schemes do not consider much human perception in GT generation, and the trained models tend to produce over-smoothed results or unpleasant artifacts. With the above considerations, we propose a human guided GT generation scheme. We first elaborately train multiple image enhancement models to improve the perceptual quality of HR images, and enable one LR image having multiple HR counterparts. Human subjects are then involved to annotate the high quality regions among the enhanced HR images as GTs, and label the regions with unpleasant artifacts as negative samples. A human guided GT image dataset with both positive and negative samples is then constructed, and a loss function is proposed to train the Real-ISR models. Experiments show that the Real-ISR models trained on our dataset can produce perceptually more realistic results with less artifacts. Dataset and codes can be found at https://github.com/ChrisDud0257/HGGT
Purpose: We propose a novel contrastive learning latent space representation for MRI datasets with partially acquired scans. We show that this latent space can be utilized for accelerated MR image reconstruction. Theory and Methods: Our novel framework, referred to as COLADA (stands for Contrastive Learning for highly accelerated MR image reconstruction), maximizes the mutual information between differently accelerated images of an MRI scan by using self-supervised contrastive learning. In other words, it attempts to "pull" the latent representations of the same scan together and "push" the latent representations of other scans away. The generated MRI latent space is subsequently utilized for MR image reconstruction and the performance was assessed in comparison to several baseline deep learning reconstruction methods. Furthermore, the quality of the proposed latent space representation was analyzed using Alignment and Uniformity. Results: COLADA comprehensively outperformed other reconstruction methods with robustness to variations in undersampling patterns, pathological abnormalities, and noise in k-space during inference. COLADA proved the high quality of reconstruction on unseen data with minimal fine-tuning. The analysis of representation quality suggests that the contrastive features produced by COLADA are optimally distributed in latent space. Conclusion: To the best of our knowledge, this is the first attempt to utilize contrastive learning on differently accelerated images for MR image reconstruction. The proposed latent space representation has practical usage due to a large number of existing partially sampled datasets. This implies the possibility of exploring self-supervised contrastive learning further to enhance the latent space of MRI for image reconstruction.
Semantic segmentation for medical 3D image stacks enables accurate volumetric reconstructions, computer-aided diagnostics and follow up treatment planning. In this work, we present a novel variant of the Unet model called the NUMSnet that transmits pixel neighborhood features across scans through nested layers to achieve accurate multi-class semantic segmentations with minimal training data. We analyze the semantic segmentation performance of the NUMSnet model in comparison with several Unet model variants to segment 3-7 regions of interest using only 10% of images for training per Lung-CT and Heart-CT volumetric image stacks. The proposed NUMSnet model achieves up to 20% improvement in segmentation recall with 4-9% improvement in Dice scores for Lung-CT stacks and 2.5-10% improvement in Dice scores for Heart-CT stacks when compared to the Unet++ model. The NUMSnet model needs to be trained by ordered images around the central scan of each volumetric stack. Propagation of image feature information from the 6 nested layers of the Unet++ model are found to have better computation and segmentation performances than propagation of all up-sampling layers in a Unet++ model. The NUMSnet model achieves comparable segmentation performances to existing works, while being trained on as low as 5\% of the training images. Also, transfer learning allows faster convergence of the NUMSnet model for multi-class semantic segmentation from pathology in Lung-CT images to cardiac segmentations in Heart-CT stacks. Thus, the proposed model can standardize multi-class semantic segmentation on a variety of volumetric image stacks with minimal training dataset. This can significantly reduce the cost, time and inter-observer variabilities associated with computer-aided detections and treatment.
Semantic image synthesis aims to generate photo realistic images given a semantic segmentation map. Despite much recent progress, training them still requires large datasets of images annotated with per-pixel label maps that are extremely tedious to obtain. To alleviate the high annotation cost, we propose a transfer method that leverages a model trained on a large source dataset to improve the learning ability on small target datasets via estimated pairwise relations between source and target classes. The class affinity matrix is introduced as a first layer to the source model to make it compatible with the target label maps, and the source model is then further finetuned for the target domain. To estimate the class affinities we consider different approaches to leverage prior knowledge: semantic segmentation on the source domain, textual label embeddings, and self-supervised vision features. We apply our approach to GAN-based and diffusion-based architectures for semantic synthesis. Our experiments show that the different ways to estimate class affinity can be effectively combined, and that our approach significantly improves over existing state-of-the-art transfer approaches for generative image models.
Recently, deep learning based approaches have shown promising results in 3D hand reconstruction from a single RGB image. These approaches can be roughly divided into model-based approaches, which are heavily dependent on the model's parameter space, and model-free approaches, which require large numbers of 3D ground truths to reduce depth ambiguity and struggle in weakly-supervised scenarios. To overcome these issues, we propose a novel probabilistic model to achieve the robustness of model-based approaches and reduced dependence on the model's parameter space of model-free approaches. The proposed probabilistic model incorporates a model-based network as a prior-net to estimate the prior probability distribution of joints and vertices. An Attention-based Mesh Vertices Uncertainty Regression (AMVUR) model is proposed to capture dependencies among vertices and the correlation between joints and mesh vertices to improve their feature representation. We further propose a learning based occlusion-aware Hand Texture Regression model to achieve high-fidelity texture reconstruction. We demonstrate the flexibility of the proposed probabilistic model to be trained in both supervised and weakly-supervised scenarios. The experimental results demonstrate our probabilistic model's state-of-the-art accuracy in 3D hand and texture reconstruction from a single image in both training schemes, including in the presence of severe occlusions.
Synthesizing novel view images from a few views is a challenging but practical problem. Existing methods often struggle with producing high-quality results or necessitate per-object optimization in such few-view settings due to the insufficient information provided. In this work, we explore leveraging the strong 2D priors in pre-trained diffusion models for synthesizing novel view images. 2D diffusion models, nevertheless, lack 3D awareness, leading to distorted image synthesis and compromising the identity. To address these problems, we propose DreamSparse, a framework that enables the frozen pre-trained diffusion model to generate geometry and identity-consistent novel view image. Specifically, DreamSparse incorporates a geometry module designed to capture 3D features from sparse views as a 3D prior. Subsequently, a spatial guidance model is introduced to convert these 3D feature maps into spatial information for the generative process. This information is then used to guide the pre-trained diffusion model, enabling it to generate geometrically consistent images without tuning it. Leveraging the strong image priors in the pre-trained diffusion models, DreamSparse is capable of synthesizing high-quality novel views for both object and scene-level images and generalising to open-set images. Experimental results demonstrate that our framework can effectively synthesize novel view images from sparse views and outperforms baselines in both trained and open-set category images. More results can be found on our project page: https://sites.google.com/view/dreamsparse-webpage.
While natural language offers a convenient shared interface for humans and robots, enabling robots to interpret and follow language commands remains a longstanding challenge in manipulation. A crucial step to realizing a performant instruction-following robot is achieving semantic manipulation, where a robot interprets language at different specificities, from high-level instructions like "Pick up the stuffed animal" to more detailed inputs like "Grab the left ear of the elephant." To tackle this, we propose Keypoints + Instructions to Execution (KITE), a two-step framework for semantic manipulation which attends to both scene semantics (distinguishing between different objects in a visual scene) and object semantics (precisely localizing different parts within an object instance). KITE first grounds an input instruction in a visual scene through 2D image keypoints, providing a highly accurate object-centric bias for downstream action inference. Provided an RGB-D scene observation, KITE then executes a learned keypoint-conditioned skill to carry out the instruction. The combined precision of keypoints and parameterized skills enables fine-grained manipulation with generalization to scene and object variations. Empirically, we demonstrate KITE in 3 real-world environments: long-horizon 6-DoF tabletop manipulation, semantic grasping, and a high-precision coffee-making task. In these settings, KITE achieves a 75%, 70%, and 71% overall success rate for instruction-following, respectively. KITE outperforms frameworks that opt for pre-trained visual language models over keypoint-based grounding, or omit skills in favor of end-to-end visuomotor control, all while being trained from fewer or comparable amounts of demonstrations. Supplementary material, datasets, code, and videos can be found on our website: http://tinyurl.com/kite-site.
Multimodal Large Language Model (MLLM) relies on the powerful LLM to perform multimodal tasks, showing amazing emergent abilities in recent studies, such as writing poems based on an image. However, it is difficult for these case studies to fully reflect the performance of MLLM, lacking a comprehensive evaluation. In this paper, we fill in this blank, presenting the first MLLM Evaluation benchmark MME. It measures both perception and cognition abilities on a total of 14 subtasks. In order to avoid data leakage that may arise from direct use of public datasets for evaluation, the annotations of instruction-answer pairs are all manually designed. The concise instruction design allows us to fairly compare MLLMs, instead of struggling in prompt engineering. Besides, with such an instruction, we can also easily carry out quantitative statistics. A total of 10 advanced MLLMs are comprehensively evaluated on our MME, which not only suggests that existing MLLMs still have a large room for improvement, but also reveals the potential directions for the subsequent model optimization.