ILLUME+: Illuminating Unified MLLM with Dual Visual Tokenization and Diffusion Refinement

We present ILLUME+ that leverages dual visual tokenization and a diffusion decoder to improve both deep semantic understanding and high-fidelity image generation. Existing unified models have struggled to simultaneously handle the three fundamental capabilities in a unified model: understanding, generation, and editing. Models like Chameleon and EMU3 utilize VQGAN for image discretization, due to the lack of deep semantic interaction, they lag behind specialist models like LLaVA in visual understanding tasks. To mitigate this, LaViT and ILLUME employ semantic encoders for tokenization, but they struggle with image editing due to poor texture preservation. Meanwhile, Janus series decouples the input and output image representation, limiting their abilities to seamlessly handle interleaved image-text understanding and generation. In contrast, ILLUME+ introduces a unified dual visual tokenizer, DualViTok, which preserves both fine-grained textures and text-aligned semantics while enabling a coarse-to-fine image representation strategy for multimodal understanding and generation. Additionally, we employ a diffusion model as the image detokenizer for enhanced generation quality and efficient super-resolution. ILLUME+ follows a continuous-input, discrete-output scheme within the unified MLLM and adopts a progressive training procedure that supports dynamic resolution across the vision tokenizer, MLLM, and diffusion decoder. This design allows for flexible and efficient context-aware image editing and generation across diverse tasks. ILLUME+ (3B) exhibits competitive performance against existing unified MLLMs and specialized models across multimodal understanding, generation, and editing benchmarks. With its strong performance, ILLUME+ provides a scalable and versatile foundation for future multimodal applications. Project Page: https://illume-unified-mllm.github.io/.

Empowering Large Language Models with 3D Situation Awareness

Driven by the great success of Large Language Models (LLMs) in the 2D image domain, their applications in 3D scene understanding has emerged as a new trend. A key difference between 3D and 2D is that the situation of an egocentric observer in 3D scenes can change, resulting in different descriptions (e.g., ''left" or ''right"). However, current LLM-based methods overlook the egocentric perspective and simply use datasets from a global viewpoint. To address this issue, we propose a novel approach to automatically generate a situation-aware dataset by leveraging the scanning trajectory during data collection and utilizing Vision-Language Models (VLMs) to produce high-quality captions and question-answer pairs. Furthermore, we introduce a situation grounding module to explicitly predict the position and orientation of observer's viewpoint, thereby enabling LLMs to ground situation description in 3D scenes. We evaluate our approach on several benchmarks, demonstrating that our method effectively enhances the 3D situational awareness of LLMs while significantly expanding existing datasets and reducing manual effort.

Sonata: Self-Supervised Learning of Reliable Point Representations

In this paper, we question whether we have a reliable self-supervised point cloud model that can be used for diverse 3D tasks via simple linear probing, even with limited data and minimal computation. We find that existing 3D self-supervised learning approaches fall short when evaluated on representation quality through linear probing. We hypothesize that this is due to what we term the "geometric shortcut", which causes representations to collapse to low-level spatial features. This challenge is unique to 3D and arises from the sparse nature of point cloud data. We address it through two key strategies: obscuring spatial information and enhancing the reliance on input features, ultimately composing a Sonata of 140k point clouds through self-distillation. Sonata is simple and intuitive, yet its learned representations are strong and reliable: zero-shot visualizations demonstrate semantic grouping, alongside strong spatial reasoning through nearest-neighbor relationships. Sonata demonstrates exceptional parameter and data efficiency, tripling linear probing accuracy (from 21.8% to 72.5%) on ScanNet and nearly doubling performance with only 1% of the data compared to previous approaches. Full fine-tuning further advances SOTA across both 3D indoor and outdoor perception tasks.

Modular Customization of Diffusion Models via Blockwise-Parameterized Low-Rank Adaptation

Recent diffusion model customization has shown impressive results in incorporating subject or style concepts with a handful of images. However, the modular composition of multiple concepts into a customized model, aimed to efficiently merge decentralized-trained concepts without influencing their identities, remains unresolved. Modular customization is essential for applications like concept stylization and multi-concept customization using concepts trained by different users. Existing post-training methods are only confined to a fixed set of concepts, and any different combinations require a new round of retraining. In contrast, instant merging methods often cause identity loss and interference of individual merged concepts and are usually limited to a small number of concepts. To address these issues, we propose BlockLoRA, an instant merging method designed to efficiently combine multiple concepts while accurately preserving individual concepts' identity. With a careful analysis of the underlying reason for interference, we develop the Randomized Output Erasure technique to minimize the interference of different customized models. Additionally, Blockwise LoRA Parameterization is proposed to reduce the identity loss during instant model merging. Extensive experiments validate the effectiveness of BlockLoRA, which can instantly merge 15 concepts of people, subjects, scenes, and styles with high fidelity.

Effective LLM Knowledge Learning via Model Generalization

Large language models (LLMs) are trained on enormous documents that contain extensive world knowledge. However, it is still not well-understood how knowledge is acquired via autoregressive pre-training. This lack of understanding greatly hinders effective knowledge learning, especially for continued pretraining on up-to-date information, as this evolving information often lacks diverse repetitions like foundational knowledge. In this paper, we focus on understanding and improving LLM knowledge learning. We found and verified that knowledge learning for LLMs can be deemed as an implicit supervised task hidden in the autoregressive pre-training objective. Our findings suggest that knowledge learning for LLMs would benefit from methods designed to improve generalization ability for supervised tasks. Based on our analysis, we propose the formatting-based data augmentation to grow in-distribution samples, which does not present the risk of altering the facts embedded in documents as text paraphrasing. We also introduce sharpness-aware minimization as an effective optimization algorithm to better improve generalization. Moreover, our analysis and method can be readily extended to instruction tuning. Extensive experiment results validate our findings and demonstrate our methods' effectiveness in both continued pre-training and instruction tuning. This paper offers new perspectives and insights to interpret and design effective strategies for LLM knowledge learning.

HERMES: A Unified Self-Driving World Model for Simultaneous 3D Scene Understanding and Generation

Driving World Models (DWMs) have become essential for autonomous driving by enabling future scene prediction. However, existing DWMs are limited to scene generation and fail to incorporate scene understanding, which involves interpreting and reasoning about the driving environment. In this paper, we present a unified Driving World Model named HERMES. We seamlessly integrate 3D scene understanding and future scene evolution (generation) through a unified framework in driving scenarios. Specifically, HERMES leverages a Bird's-Eye View (BEV) representation to consolidate multi-view spatial information while preserving geometric relationships and interactions. We also introduce world queries, which incorporate world knowledge into BEV features via causal attention in the Large Language Model (LLM), enabling contextual enrichment for understanding and generation tasks. We conduct comprehensive studies on nuScenes and OmniDrive-nuScenes datasets to validate the effectiveness of our method. HERMES achieves state-of-the-art performance, reducing generation error by 32.4% and improving understanding metrics such as CIDEr by 8.0%. The model and code will be publicly released at https://github.com/LMD0311/HERMES.

DiffDoctor: Diagnosing Image Diffusion Models Before Treating

In spite of the recent progress, image diffusion models still produce artifacts. A common solution is to refine an established model with a quality assessment system, which generally rates an image in its entirety. In this work, we believe problem-solving starts with identification, yielding the request that the model should be aware of not just the presence of defects in an image, but their specific locations. Motivated by this, we propose DiffDoctor, a two-stage pipeline to assist image diffusion models in generating fewer artifacts. Concretely, the first stage targets developing a robust artifact detector, for which we collect a dataset of over 1M flawed synthesized images and set up an efficient human-in-the-loop annotation process, incorporating a carefully designed class-balance strategy. The learned artifact detector is then involved in the second stage to tune the diffusion model through assigning a per-pixel confidence map for each synthesis. Extensive experiments on text-to-image diffusion models demonstrate the effectiveness of our artifact detector as well as the soundness of our diagnose-then-treat design.

VideoAnydoor: High-fidelity Video Object Insertion with Precise Motion Control

Despite significant advancements in video generation, inserting a given object into videos remains a challenging task. The difficulty lies in preserving the appearance details of the reference object and accurately modeling coherent motions at the same time. In this paper, we propose VideoAnydoor, a zero-shot video object insertion framework with high-fidelity detail preservation and precise motion control. Starting from a text-to-video model, we utilize an ID extractor to inject the global identity and leverage a box sequence to control the overall motion. To preserve the detailed appearance and meanwhile support fine-grained motion control, we design a pixel warper. It takes the reference image with arbitrary key-points and the corresponding key-point trajectories as inputs. It warps the pixel details according to the trajectories and fuses the warped features with the diffusion U-Net, thus improving detail preservation and supporting users in manipulating the motion trajectories. In addition, we propose a training strategy involving both videos and static images with a weighted loss to enhance insertion quality. VideoAnydoor demonstrates significant superiority over existing methods and naturally supports various downstream applications (e.g., talking head generation, video virtual try-on, multi-region editing) without task-specific fine-tuning.

DreamMask: Boosting Open-vocabulary Panoptic Segmentation with Synthetic Data

Open-vocabulary panoptic segmentation has received significant attention due to its applicability in the real world. Despite claims of robust generalization, we find that the advancements of previous works are attributed mainly on trained categories, exposing a lack of generalization to novel classes. In this paper, we explore boosting existing models from a data-centric perspective. We propose DreamMask, which systematically explores how to generate training data in the open-vocabulary setting, and how to train the model with both real and synthetic data. For the first part, we propose an automatic data generation pipeline with off-the-shelf models. We propose crucial designs for vocabulary expansion, layout arrangement, data filtering, etc. Equipped with these techniques, our generated data could significantly outperform the manually collected web data. To train the model with generated data, a synthetic-real alignment loss is designed to bridge the representation gap, bringing noticeable improvements across multiple benchmarks. In general, DreamMask significantly simplifies the collection of large-scale training data, serving as a plug-and-play enhancement for existing methods. For instance, when trained on COCO and tested on ADE20K, the model equipped with DreamMask outperforms the previous state-of-the-art by a substantial margin of 2.1% mIoU.

GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models

In recent years, 2D Vision-Language Models (VLMs) have made significant strides in image-text understanding tasks. However, their performance in 3D spatial comprehension, which is critical for embodied intelligence, remains limited. Recent advances have leveraged 3D point clouds and multi-view images as inputs, yielding promising results. However, we propose exploring a purely vision-based solution inspired by human perception, which merely relies on visual cues for 3D spatial understanding. This paper empirically investigates the limitations of VLMs in 3D spatial knowledge, revealing that their primary shortcoming lies in the lack of global-local correspondence between the scene and individual frames. To address this, we introduce GPT4Scene, a novel visual prompting paradigm in VLM training and inference that helps build the global-local relationship, significantly improving the 3D spatial understanding of indoor scenes. Specifically, GPT4Scene constructs a 3D Bird's Eye View (BEV) image from the video and marks consistent object IDs across both frames and the BEV image. The model then inputs the concatenated BEV image and video frames with markers. In zero-shot evaluations, GPT4Scene improves performance over closed-source VLMs like GPT-4o. Additionally, we prepare a processed video dataset consisting of 165K text annotation to fine-tune open-source VLMs, achieving state-of-the-art performance on all 3D understanding tasks. Surprisingly, after training with the GPT4Scene paradigm, VLMs consistently improve during inference, even without visual prompting and BEV image as explicit correspondence. It demonstrates that the proposed paradigm helps VLMs develop an intrinsic ability to understand 3D scenes, which paves the way for a noninvasive approach to extending pre-trained VLMs for 3D scene understanding.