Haotian Liu

Zhiqing Sun, Sheng Shen, Shengcao Cao et al. · berkeley, cmu

Large Multimodal Models (LMM) are built across modalities and the misalignment between two modalities can result in "hallucination", generating textual outputs that are not grounded by the multimodal information in context. To address the multimodal misalignment issue, we adapt the Reinforcement Learning from Human Feedback (RLHF) from the text domain to the task of vision-language alignment, where human annotators are asked to compare two responses and pinpoint the more hallucinated one, and the vision-language model is trained to maximize the simulated human rewards. We propose a new alignment algorithm called Factually Augmented RLHF that augments the reward model with additional factual information such as image captions and ground-truth multi-choice options, which alleviates the reward hacking phenomenon in RLHF and further improves the performance. We also enhance the GPT-4-generated training data (for vision instruction tuning) with previously available human-written image-text pairs to improve the general capabilities of our model. To evaluate the proposed approach in real-world scenarios, we develop a new evaluation benchmark MMHAL-BENCH with a special focus on penalizing hallucinations. As the first LMM trained with RLHF, our approach achieves remarkable improvement on the LLaVA-Bench dataset with the 94% performance level of the text-only GPT-4 (while previous best methods can only achieve the 87% level), and an improvement by 60% on MMHAL-BENCH over other baselines. We opensource our code, model, data at https://llava-rlhf.github.io.

Shilong Liu, Hao Cheng, Haotian Liu et al. · microsoft-research

LLaVA-Plus is a general-purpose multimodal assistant that expands the capabilities of large multimodal models. It maintains a skill repository of pre-trained vision and vision-language models and can activate relevant tools based on users' inputs to fulfill real-world tasks. LLaVA-Plus is trained on multimodal instruction-following data to acquire the ability to use tools, covering visual understanding, generation, external knowledge retrieval, and compositions. Empirical results show that LLaVA-Plus outperforms LLaVA in existing capabilities and exhibits new ones. It is distinct in that the image query is directly grounded and actively engaged throughout the entire human-AI interaction sessions, significantly improving tool use performance and enabling new scenarios.

Haotian Liu, Chunyuan Li, Yuheng Li et al.

Large multimodal models (LMM) have recently shown encouraging progress with visual instruction tuning. In this note, we show that the fully-connected vision-language cross-modal connector in LLaVA is surprisingly powerful and data-efficient. With simple modifications to LLaVA, namely, using CLIP-ViT-L-336px with an MLP projection and adding academic-task-oriented VQA data with simple response formatting prompts, we establish stronger baselines that achieve state-of-the-art across 11 benchmarks. Our final 13B checkpoint uses merely 1.2M publicly available data, and finishes full training in ~1 day on a single 8-A100 node. We hope this can make state-of-the-art LMM research more accessible. Code and model will be publicly available.

Fangzhou Lin, Haotian Liu, Haoying Zhou et al.

3D point clouds enhanced the robot's ability to perceive the geometrical information of the environments, making it possible for many downstream tasks such as grasp pose detection and scene understanding. The performance of these tasks, though, heavily relies on the quality of data input, as incomplete can lead to poor results and failure cases. Recent training loss functions designed for deep learning-based point cloud completion, such as Chamfer distance (CD) and its variants (\eg HyperCD ), imply a good gradient weighting scheme can significantly boost performance. However, these CD-based loss functions usually require data-related parameter tuning, which can be time-consuming for data-extensive tasks. To address this issue, we aim to find a family of weighted training losses ({\em weighted CD}) that requires no parameter tuning. To this end, we propose a search scheme, {\em Loss Distillation via Gradient Matching}, to find good candidate loss functions by mimicking the learning behavior in backpropagation between HyperCD and weighted CD. Once this is done, we propose a novel bilevel optimization formula to train the backbone network based on the weighted CD loss. We observe that: (1) with proper weighted functions, the weighted CD can always achieve similar performance to HyperCD, and (2) the Landau weighted CD, namely {\em Landau CD}, can outperform HyperCD for point cloud completion and lead to new state-of-the-art results on several benchmark datasets. {\it Our demo code is available at \url{https://github.com/Zhang-VISLab/IROS2024-LossDistillationWeightedCD}.}

Fangzhou Lin, Songlin Hou, Haotian Liu et al.

Chamfer Distance (CD) is widely used as a metric to quantify difference between two point clouds. In point cloud completion, Chamfer Distance (CD) is typically used as a loss function in deep learning frameworks. However, it is generally acknowledged within the field that Chamfer Distance (CD) is vulnerable to the presence of outliers, which can consequently lead to the convergence on suboptimal models. In divergence from the existing literature, which largely concentrates on resolving such concerns in the realm of Euclidean space, we put forth a notably uncomplicated yet potent metric specifically designed for point cloud completion tasks: {Hyperbolic Chamfer Distance (HyperCD)}. This metric conducts Chamfer Distance computations within the parameters of hyperbolic space. During the backpropagation process, HyperCD systematically allocates greater weight to matched point pairs exhibiting reduced Euclidean distances. This mechanism facilitates the preservation of accurate point pair matches while permitting the incremental adjustment of suboptimal matches, thereby contributing to enhanced point cloud completion outcomes. Moreover, measure the shape dissimilarity is not solely work for point cloud completion task, we further explore its applications in other generative related tasks, including single image reconstruction from point cloud, and upsampling. We demonstrate state-of-the-art performance on the point cloud completion benchmark datasets, PCN, ShapeNet-55, and ShapeNet-34, and show from visualization that HyperCD can significantly improve the surface smoothness, we also provide the provide experimental results beyond completion task.

Thao Nguyen, Haotian Liu, Yuheng Li et al.

Large Multimodal Models (LMMs) have shown remarkable capabilities across a variety of tasks (e.g., image captioning, visual question answering). While broad, their knowledge remains generic (e.g., recognizing a dog), and they are unable to handle personalized subjects (e.g., recognizing a user's pet dog). Human reasoning, in contrast, typically operates within the context of specific subjects in our surroundings. For example, one might ask, "What should I buy for my dog's birthday?"; as opposed to a generic inquiry about "What should I buy for a dog's birthday?". Similarly, when looking at a friend's image, the interest lies in seeing their activities (e.g., "my friend is holding a cat"), rather than merely observing generic human actions (e.g., "a man is holding a cat"). In this paper, we introduce the novel task of personalizing LMMs, so that they can have conversations about a specific subject. We propose Yo'LLaVA, which learns to embed a personalized subject into a set of latent tokens given a handful of example images of the subject. Our qualitative and quantitative analyses reveal that Yo'LLaVA can learn the concept more efficiently using fewer tokens and more effectively encode the visual attributes compared to strong prompting baselines (e.g., LLaVA).

Thao Nguyen, Utkarsh Ojha, Yuheng Li et al.

In recent years, image editing has advanced remarkably. With increased human control, it is now possible to edit an image in a plethora of ways; from specifying in text what we want to change, to straight up dragging the contents of the image in an interactive point-based manner. However, most of the focus has remained on editing single images at a time. Whether and how we can simultaneously edit large batches of images has remained understudied. With the goal of minimizing human supervision in the editing process, this paper presents a novel method for interactive batch image editing using StyleGAN as the medium. Given an edit specified by users in an example image (e.g., make the face frontal), our method can automatically transfer that edit to other test images, so that regardless of their initial state (pose), they all arrive at the same final state (e.g., all facing front). Extensive experiments demonstrate that edits performed using our method have similar visual quality to existing single-image-editing methods, while having more visual consistency and saving significant time and human effort.