From Visuals to Vocabulary: Establishing Equivalence Between Image and Text Token Through Autoregressive Pre-training in MLLMs

While MLLMs perform well on perceptual tasks, they lack precise multimodal alignment, limiting performance. To address this challenge, we propose Vision Dynamic Embedding-Guided Pretraining (VDEP), a hybrid autoregressive training paradigm for MLLMs. Utilizing dynamic embeddings from the MLP following the visual encoder, this approach supervises image hidden states and integrates image tokens into autoregressive training. Existing MLLMs primarily focused on recovering information from textual inputs, often neglecting the effective processing of image data. In contrast, the key improvement of this work is the reinterpretation of multimodal alignment as a process of recovering information from input data, with particular emphasis on reconstructing detailed visual features.The proposed method seamlessly integrates into standard models without architectural changes. Experiments on 13 benchmarks show VDEP outperforms baselines, surpassing existing methods.

Test Time Training for 4D Medical Image Interpolation

4D medical image interpolation is essential for improving temporal resolution and diagnostic precision in clinical applications. Previous works ignore the problem of distribution shifts, resulting in poor generalization under different distribution. A natural solution would be to adapt the model to a new test distribution, but this cannot be done if the test input comes without a ground truth label. In this paper, we propose a novel test time training framework which uses self-supervision to adapt the model to a new distribution without requiring any labels. Indeed, before performing frame interpolation on each test video, the model is trained on the same instance using a self-supervised task, such as rotation prediction or image reconstruction. We conduct experiments on two publicly available 4D medical image interpolation datasets, Cardiac and 4D-Lung. The experimental results show that the proposed method achieves significant performance across various evaluation metrics on both datasets. It achieves higher peak signal-to-noise ratio values, 33.73dB on Cardiac and 34.02dB on 4D-Lung. Our method not only advances 4D medical image interpolation but also provides a template for domain adaptation in other fields such as image segmentation and image registration.

GPS as a Control Signal for Image Generation

We show that the GPS tags contained in photo metadata provide a useful control signal for image generation. We train GPS-to-image models and use them for tasks that require a fine-grained understanding of how images vary within a city. In particular, we train a diffusion model to generate images conditioned on both GPS and text. The learned model generates images that capture the distinctive appearance of different neighborhoods, parks, and landmarks. We also extract 3D models from 2D GPS-to-image models through score distillation sampling, using GPS conditioning to constrain the appearance of the reconstruction from each viewpoint. Our evaluations suggest that our GPS-conditioned models successfully learn to generate images that vary based on location, and that GPS conditioning improves estimated 3D structure.

Advancing General Multimodal Capability of Vision-language Models with Pyramid-descent Visual Position Encoding

Vision-language Models (VLMs) have shown remarkable capabilities in advancing general artificial intelligence, yet the irrational encoding of visual positions persists in inhibiting the models' comprehensive perception performance across different levels of granularity. In this work, we propose Pyramid-descent Visual Position Encoding (PyPE), a novel approach designed to enhance the perception of visual tokens within VLMs. By assigning visual position indexes from the periphery to the center and expanding the central receptive field incrementally, PyPE addresses the limitations of traditional raster-scan methods and mitigates the long-term decay effects induced by Rotary Position Embedding (RoPE). Our method reduces the relative distance between interrelated visual elements and instruction tokens, promoting a more rational allocation of attention weights and allowing for a multi-granularity perception of visual elements and countering the over-reliance on anchor tokens. Extensive experimental evaluations demonstrate that PyPE consistently improves the general capabilities of VLMs across various sizes. Code is available at https://github.com/SakuraTroyChen/PyPE.

Hadamard Attention Recurrent Transformer: A Strong Baseline for Stereo Matching Transformer

In light of the advancements in transformer technology, extant research posits the construction of stereo transformers as a potential solution to the binocular stereo matching challenge. However, constrained by the low-rank bottleneck and quadratic complexity of attention mechanisms, stereo transformers still fail to demonstrate sufficient nonlinear expressiveness within a reasonable inference time. The lack of focus on key homonymous points renders the representations of such methods vulnerable to challenging conditions, including reflections and weak textures. Furthermore, a slow computing speed is not conducive to the application. To overcome these difficulties, we present the \textbf{H}adamard \textbf{A}ttention \textbf{R}ecurrent Stereo \textbf{T}ransformer (HART) that incorporates the following components: 1) For faster inference, we present a Hadamard product paradigm for the attention mechanism, achieving linear computational complexity. 2) We designed a Dense Attention Kernel (DAK) to amplify the differences between relevant and irrelevant feature responses. This allows HART to focus on important details. DAK also converts zero elements to non-zero elements to mitigate the reduced expressiveness caused by the low-rank bottleneck. 3) To compensate for the spatial and channel interaction missing in the Hadamard product, we propose MKOI to capture both global and local information through the interleaving of large and small kernel convolutions. Experimental results demonstrate the effectiveness of our HART. In reflective area, HART ranked \textbf{1st} on the KITTI 2012 benchmark among all published methods at the time of submission. Code is available at \url{https://github.com/ZYangChen/HART}.

Meta Curvature-Aware Minimization for Domain Generalization

Domain generalization (DG) aims to enhance the ability of models trained on source domains to generalize effectively to unseen domains. Recently, Sharpness-Aware Minimization (SAM) has shown promise in this area by reducing the sharpness of the loss landscape to obtain more generalized models. However, SAM and its variants sometimes fail to guide the model toward a flat minimum, and their training processes exhibit limitations, hindering further improvements in model generalization. In this paper, we first propose an improved model training process aimed at encouraging the model to converge to a flat minima. To achieve this, we design a curvature metric that has a minimal effect when the model is far from convergence but becomes increasingly influential in indicating the curvature of the minima as the model approaches a local minimum. Then we derive a novel algorithm from this metric, called Meta Curvature-Aware Minimization (MeCAM), to minimize the curvature around the local minima. Specifically, the optimization objective of MeCAM simultaneously minimizes the regular training loss, the surrogate gap of SAM, and the surrogate gap of meta-learning. We provide theoretical analysis on MeCAM's generalization error and convergence rate, and demonstrate its superiority over existing DG methods through extensive experiments on five benchmark DG datasets, including PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet. Code will be available on GitHub.

Video-Guided Foley Sound Generation with Multimodal Controls

Generating sound effects for videos often requires creating artistic sound effects that diverge significantly from real-life sources and flexible control in the sound design. To address this problem, we introduce MultiFoley, a model designed for video-guided sound generation that supports multimodal conditioning through text, audio, and video. Given a silent video and a text prompt, MultiFoley allows users to create clean sounds (e.g., skateboard wheels spinning without wind noise) or more whimsical sounds (e.g., making a lion's roar sound like a cat's meow). MultiFoley also allows users to choose reference audio from sound effects (SFX) libraries or partial videos for conditioning. A key novelty of our model lies in its joint training on both internet video datasets with low-quality audio and professional SFX recordings, enabling high-quality, full-bandwidth (48kHz) audio generation. Through automated evaluations and human studies, we demonstrate that MultiFoley successfully generates synchronized high-quality sounds across varied conditional inputs and outperforms existing methods. Please see our project page for video results: https://ificl.github.io/MultiFoley/

Motif Channel Opened in a White-Box: Stereo Matching via Motif Correlation Graph

Real-world applications of stereo matching, such as autonomous driving, place stringent demands on both safety and accuracy. However, learning-based stereo matching methods inherently suffer from the loss of geometric structures in certain feature channels, creating a bottleneck in achieving precise detail matching. Additionally, these methods lack interpretability due to the black-box nature of deep learning. In this paper, we propose MoCha-V2, a novel learning-based paradigm for stereo matching. MoCha-V2 introduces the Motif Correlation Graph (MCG) to capture recurring textures, which are referred to as ``motifs" within feature channels. These motifs reconstruct geometric structures and are learned in a more interpretable way. Subsequently, we integrate features from multiple frequency domains through wavelet inverse transformation. The resulting motif features are utilized to restore geometric structures in the stereo matching process. Experimental results demonstrate the effectiveness of MoCha-V2. MoCha-V2 achieved 1st place on the Middlebury benchmark at the time of its release. Code is available at https://github.com/ZYangChen/MoCha-Stereo.

CoSAM: Self-Correcting SAM for Domain Generalization in 2D Medical Image Segmentation

Medical images often exhibit distribution shifts due to variations in imaging protocols and scanners across different medical centers. Domain Generalization (DG) methods aim to train models on source domains that can generalize to unseen target domains. Recently, the segment anything model (SAM) has demonstrated strong generalization capabilities due to its prompt-based design, and has gained significant attention in image segmentation tasks. Existing SAM-based approaches attempt to address the need for manual prompts by introducing prompt generators that automatically generate these prompts. However, we argue that auto-generated prompts may not be sufficiently accurate under distribution shifts, potentially leading to incorrect predictions that still require manual verification and correction by clinicians. To address this challenge, we propose a method for 2D medical image segmentation called Self-Correcting SAM (CoSAM). Our approach begins by generating coarse masks using SAM in a prompt-free manner, providing prior prompts for the subsequent stages, and eliminating the need for prompt generators. To automatically refine these coarse masks, we introduce a generalized error decoder that simulates the correction process typically performed by clinicians. Furthermore, we generate diverse prompts as feedback based on the corrected masks, which are used to iteratively refine the predictions within a self-correcting loop, enhancing the generalization performance of our model. Extensive experiments on two medical image segmentation benchmarks across multiple scenarios demonstrate the superiority of CoSAM over state-of-the-art SAM-based methods.

Touchstone Benchmark: Are We on the Right Way for Evaluating AI Algorithms for Medical Segmentation?

How can we test AI performance? This question seems trivial, but it isn't. Standard benchmarks often have problems such as in-distribution and small-size test sets, oversimplified metrics, unfair comparisons, and short-term outcome pressure. As a consequence, good performance on standard benchmarks does not guarantee success in real-world scenarios. To address these problems, we present Touchstone, a large-scale collaborative segmentation benchmark of 9 types of abdominal organs. This benchmark is based on 5,195 training CT scans from 76 hospitals around the world and 5,903 testing CT scans from 11 additional hospitals. This diverse test set enhances the statistical significance of benchmark results and rigorously evaluates AI algorithms across various out-of-distribution scenarios. We invited 14 inventors of 19 AI algorithms to train their algorithms, while our team, as a third party, independently evaluated these algorithms on three test sets. In addition, we also evaluated pre-existing AI frameworks--which, differing from algorithms, are more flexible and can support different algorithms--including MONAI from NVIDIA, nnU-Net from DKFZ, and numerous other open-source frameworks. We are committed to expanding this benchmark to encourage more innovation of AI algorithms for the medical domain.