DiffDis: Empowering Generative Diffusion Model with Cross-Modal Discrimination Capability

Recently, large-scale diffusion models, e.g., Stable diffusion and DallE2, have shown remarkable results on image synthesis. On the other hand, large-scale cross-modal pre-trained models (e.g., CLIP, ALIGN, and FILIP) are competent for various downstream tasks by learning to align vision and language embeddings. In this paper, we explore the possibility of jointly modeling generation and discrimination. Specifically, we propose DiffDis to unify the cross-modal generative and discriminative pretraining into one single framework under the diffusion process. DiffDis first formulates the image-text discriminative problem as a generative diffusion process of the text embedding from the text encoder conditioned on the image. Then, we propose a novel dual-stream network architecture, which fuses the noisy text embedding with the knowledge of latent images from different scales for image-text discriminative learning. Moreover, the generative and discriminative tasks can efficiently share the image-branch network structure in the multi-modality model. Benefiting from diffusion-based unified training, DiffDis achieves both better generation ability and cross-modal semantic alignment in one architecture. Experimental results show that DiffDis outperforms single-task models on both the image generation and the image-text discriminative tasks, e.g., 1.65% improvement on average accuracy of zero-shot classification over 12 datasets and 2.42 improvement on FID of zero-shot image synthesis.

CTP: Towards Vision-Language Continual Pretraining via Compatible Momentum Contrast and Topology Preservation

Vision-Language Pretraining (VLP) has shown impressive results on diverse downstream tasks by offline training on large-scale datasets. Regarding the growing nature of real-world data, such an offline training paradigm on ever-expanding data is unsustainable, because models lack the continual learning ability to accumulate knowledge constantly. However, most continual learning studies are limited to uni-modal classification and existing multi-modal datasets cannot simulate continual non-stationary data stream scenarios. To support the study of Vision-Language Continual Pretraining (VLCP), we first contribute a comprehensive and unified benchmark dataset P9D which contains over one million product image-text pairs from 9 industries. The data from each industry as an independent task supports continual learning and conforms to the real-world long-tail nature to simulate pretraining on web data. We comprehensively study the characteristics and challenges of VLCP, and propose a new algorithm: Compatible momentum contrast with Topology Preservation, dubbed CTP. The compatible momentum model absorbs the knowledge of the current and previous-task models to flexibly update the modal feature. Moreover, Topology Preservation transfers the knowledge of embedding across tasks while preserving the flexibility of feature adjustment. The experimental results demonstrate our method not only achieves superior performance compared with other baselines but also does not bring an expensive training burden. Dataset and codes are available at https://github.com/KevinLight831/CTP.

LAW-Diffusion: Complex Scene Generation by Diffusion with Layouts

Thanks to the rapid development of diffusion models, unprecedented progress has been witnessed in image synthesis. Prior works mostly rely on pre-trained linguistic models, but a text is often too abstract to properly specify all the spatial properties of an image, e.g., the layout configuration of a scene, leading to the sub-optimal results of complex scene generation. In this paper, we achieve accurate complex scene generation by proposing a semantically controllable Layout-AWare diffusion model, termed LAW-Diffusion. Distinct from the previous Layout-to-Image generation (L2I) methods that only explore category-aware relationships, LAW-Diffusion introduces a spatial dependency parser to encode the location-aware semantic coherence across objects as a layout embedding and produces a scene with perceptually harmonious object styles and contextual relations. To be specific, we delicately instantiate each object's regional semantics as an object region map and leverage a location-aware cross-object attention module to capture the spatial dependencies among those disentangled representations. We further propose an adaptive guidance schedule for our layout guidance to mitigate the trade-off between the regional semantic alignment and the texture fidelity of generated objects. Moreover, LAW-Diffusion allows for instance reconfiguration while maintaining the other regions in a synthesized image by introducing a layout-aware latent grafting mechanism to recompose its local regional semantics. To better verify the plausibility of generated scenes, we propose a new evaluation metric for the L2I task, dubbed Scene Relation Score (SRS) to measure how the images preserve the rational and harmonious relations among contextual objects. Comprehensive experiments demonstrate that our LAW-Diffusion yields the state-of-the-art generative performance, especially with coherent object relations.

MixReorg: Cross-Modal Mixed Patch Reorganization is a Good Mask Learner for Open-World Semantic Segmentation

Recently, semantic segmentation models trained with image-level text supervision have shown promising results in challenging open-world scenarios. However, these models still face difficulties in learning fine-grained semantic alignment at the pixel level and predicting accurate object masks. To address this issue, we propose MixReorg, a novel and straightforward pre-training paradigm for semantic segmentation that enhances a model's ability to reorganize patches mixed across images, exploring both local visual relevance and global semantic coherence. Our approach involves generating fine-grained patch-text pairs data by mixing image patches while preserving the correspondence between patches and text. The model is then trained to minimize the segmentation loss of the mixed images and the two contrastive losses of the original and restored features. With MixReorg as a mask learner, conventional text-supervised semantic segmentation models can achieve highly generalizable pixel-semantic alignment ability, which is crucial for open-world segmentation. After training with large-scale image-text data, MixReorg models can be applied directly to segment visual objects of arbitrary categories, without the need for further fine-tuning. Our proposed framework demonstrates strong performance on popular zero-shot semantic segmentation benchmarks, outperforming GroupViT by significant margins of 5.0%, 6.2%, 2.5%, and 3.4% mIoU on PASCAL VOC2012, PASCAL Context, MS COCO, and ADE20K, respectively.

FULLER: Unified Multi-modality Multi-task 3D Perception via Multi-level Gradient Calibration

Multi-modality fusion and multi-task learning are becoming trendy in 3D autonomous driving scenario, considering robust prediction and computation budget. However, naively extending the existing framework to the domain of multi-modality multi-task learning remains ineffective and even poisonous due to the notorious modality bias and task conflict. Previous works manually coordinate the learning framework with empirical knowledge, which may lead to sub-optima. To mitigate the issue, we propose a novel yet simple multi-level gradient calibration learning framework across tasks and modalities during optimization. Specifically, the gradients, produced by the task heads and used to update the shared backbone, will be calibrated at the backbone's last layer to alleviate the task conflict. Before the calibrated gradients are further propagated to the modality branches of the backbone, their magnitudes will be calibrated again to the same level, ensuring the downstream tasks pay balanced attention to different modalities. Experiments on large-scale benchmark nuScenes demonstrate the effectiveness of the proposed method, eg, an absolute 14.4% mIoU improvement on map segmentation and 1.4% mAP improvement on 3D detection, advancing the application of 3D autonomous driving in the domain of multi-modality fusion and multi-task learning. We also discuss the links between modalities and tasks.

Fashion Matrix: Editing Photos by Just Talking

The utilization of Large Language Models (LLMs) for the construction of AI systems has garnered significant attention across diverse fields. The extension of LLMs to the domain of fashion holds substantial commercial potential but also inherent challenges due to the intricate semantic interactions in fashion-related generation. To address this issue, we developed a hierarchical AI system called Fashion Matrix dedicated to editing photos by just talking. This system facilitates diverse prompt-driven tasks, encompassing garment or accessory replacement, recoloring, addition, and removal. Specifically, Fashion Matrix employs LLM as its foundational support and engages in iterative interactions with users. It employs a range of Semantic Segmentation Models (e.g., Grounded-SAM, MattingAnything, etc.) to delineate the specific editing masks based on user instructions. Subsequently, Visual Foundation Models (e.g., Stable Diffusion, ControlNet, etc.) are leveraged to generate edited images from text prompts and masks, thereby facilitating the automation of fashion editing processes. Experiments demonstrate the outstanding ability of Fashion Matrix to explores the collaborative potential of functionally diverse pre-trained models in the domain of fashion editing.

RM-PRT: Realistic Robotic Manipulation Simulator and Benchmark with Progressive Reasoning Tasks

Recently, the advent of pre-trained large-scale language models (LLMs) like ChatGPT and GPT-4 have significantly advanced the machine's natural language understanding capabilities. This breakthrough has allowed us to seamlessly integrate these open-source LLMs into a unified robot simulator environment to help robots accurately understand and execute human natural language instructions. To this end, in this work, we introduce a realistic robotic manipulation simulator and build a Robotic Manipulation with Progressive Reasoning Tasks (RM-PRT) benchmark on this basis. Specifically, the RM-PRT benchmark builds a new high-fidelity digital twin scene based on Unreal Engine 5, which includes 782 categories, 2023 objects, and 15K natural language instructions generated by ChatGPT for a detailed evaluation of robot manipulation. We propose a general pipeline for the RM-PRT benchmark that takes as input multimodal prompts containing natural language instructions and automatically outputs actions containing the movement and position transitions. We set four natural language understanding tasks with progressive reasoning levels and evaluate the robot's ability to understand natural language instructions in two modes of adsorption and grasping. In addition, we also conduct a comprehensive analysis and comparison of the differences and advantages of 10 different LLMs in instruction understanding and generation quality. We hope the new simulator and benchmark will facilitate future research on language-guided robotic manipulation. Project website: https://necolizer.github.io/RM-PRT/ .

MO-VLN: A Multi-Task Benchmark for Open-set Zero-Shot Vision-and-Language Navigation

Given a natural language, a general robot has to comprehend the instruction and find the target object or location based on visual observations even in unexplored environments. Most agents rely on massive diverse training data to achieve better generalization, which requires expensive labor. These agents often focus on common objects and fewer tasks, thus are not intelligent enough to handle different types of instructions. To facilitate research in open-set vision-and-language navigation, we propose a benchmark named MO-VLN, aiming at testing the effectiveness and generalization of the agent in the multi-task setting. First, we develop a 3D simulator rendered by realistic scenarios using Unreal Engine 5, containing more realistic lights and details. The simulator contains three scenes, i.e., cafe, restaurant, and nursing house, of high value in the industry. Besides, our simulator involves multiple uncommon objects, such as takeaway cup and medical adhesive tape, which are more complicated compared with existing environments. Inspired by the recent success of large language models (e.g., ChatGPT, Vicuna), we construct diverse high-quality data of instruction type without human annotation. Our benchmark MO-VLN provides four tasks: 1) goal-conditioned navigation given a specific object category (e.g., "fork"); 2) goal-conditioned navigation given simple instructions (e.g., "Search for and move towards a tennis ball"); 3) step-by-step instruction following; 4) finding abstract object based on high-level instruction (e.g., "I am thirsty").

Boosting Text-to-Image Diffusion Models with Fine-Grained Semantic Rewards

Recent advances in text-to-image diffusion models have achieved remarkable success in generating high-quality, realistic images from given text prompts. However, previous methods fail to perform accurate modality alignment between text concepts and generated images due to the lack of fine-level semantic guidance that successfully diagnoses the modality discrepancy. In this paper, we propose FineRewards to improve the alignment between text and images in text-to-image diffusion models by introducing two new fine-grained semantic rewards: the caption reward and the Semantic Segment Anything (SAM) reward. From the global semantic view, the caption reward generates a corresponding detailed caption that depicts all important contents in the synthetic image via a BLIP-2 model and then calculates the reward score by measuring the similarity between the generated caption and the given prompt. From the local semantic view, the SAM reward segments the generated images into local parts with category labels, and scores the segmented parts by measuring the likelihood of each category appearing in the prompted scene via a large language model, i.e., Vicuna-7B. Additionally, we adopt an assemble reward-ranked learning strategy to enable the integration of multiple reward functions to jointly guide the model training. Adapting results of text-to-image models on the MS-COCO benchmark show that the proposed semantic reward outperforms other baseline reward functions with a considerable margin on both visual quality and semantic similarity with the input prompt. Moreover, by adopting the assemble reward-ranked learning strategy, we further demonstrate that model performance is further improved when adapting under the unifying of the proposed semantic reward with the current image rewards.

UniDiff: Advancing Vision-Language Models with Generative and Discriminative Learning

Recent advances in vision-language pre-training have enabled machines to perform better in multimodal object discrimination (e.g., image-text semantic alignment) and image synthesis (e.g., text-to-image generation). On the other hand, fine-tuning pre-trained models with discriminative or generative capabilities such as CLIP and Stable Diffusion on domain-specific datasets has shown to be effective in various tasks by adapting to specific domains. However, few studies have explored the possibility of learning both discriminative and generative capabilities and leveraging their synergistic effects to create a powerful and personalized multimodal model during fine-tuning. This paper presents UniDiff, a unified multi-modal model that integrates image-text contrastive learning (ITC), text-conditioned image synthesis learning (IS), and reciprocal semantic consistency modeling (RSC). UniDiff effectively learns aligned semantics and mitigates the issue of semantic collapse during fine-tuning on small datasets by leveraging RSC on visual features from CLIP and diffusion models, without altering the pre-trained model's basic architecture. UniDiff demonstrates versatility in both multi-modal understanding and generative tasks. Experimental results on three datasets (Fashion-man, Fashion-woman, and E-commercial Product) showcase substantial enhancements in vision-language retrieval and text-to-image generation, illustrating the advantages of combining discriminative and generative fine-tuning. The proposed UniDiff model establishes a robust pipeline for personalized modeling and serves as a benchmark for future comparisons in the field.