One-Minute Video Generation with Test-Time Training

Transformers today still struggle to generate one-minute videos because self-attention layers are inefficient for long context. Alternatives such as Mamba layers struggle with complex multi-scene stories because their hidden states are less expressive. We experiment with Test-Time Training (TTT) layers, whose hidden states themselves can be neural networks, therefore more expressive. Adding TTT layers into a pre-trained Transformer enables it to generate one-minute videos from text storyboards. For proof of concept, we curate a dataset based on Tom and Jerry cartoons. Compared to baselines such as Mamba~2, Gated DeltaNet, and sliding-window attention layers, TTT layers generate much more coherent videos that tell complex stories, leading by 34 Elo points in a human evaluation of 100 videos per method. Although promising, results still contain artifacts, likely due to the limited capability of the pre-trained 5B model. The efficiency of our implementation can also be improved. We have only experimented with one-minute videos due to resource constraints, but the approach can be extended to longer videos and more complex stories. Sample videos, code and annotations are available at: https://test-time-training.github.io/video-dit

Visual Acoustic Fields

Objects produce different sounds when hit, and humans can intuitively infer how an object might sound based on its appearance and material properties. Inspired by this intuition, we propose Visual Acoustic Fields, a framework that bridges hitting sounds and visual signals within a 3D space using 3D Gaussian Splatting (3DGS). Our approach features two key modules: sound generation and sound localization. The sound generation module leverages a conditional diffusion model, which takes multiscale features rendered from a feature-augmented 3DGS to generate realistic hitting sounds. Meanwhile, the sound localization module enables querying the 3D scene, represented by the feature-augmented 3DGS, to localize hitting positions based on the sound sources. To support this framework, we introduce a novel pipeline for collecting scene-level visual-sound sample pairs, achieving alignment between captured images, impact locations, and corresponding sounds. To the best of our knowledge, this is the first dataset to connect visual and acoustic signals in a 3D context. Extensive experiments on our dataset demonstrate the effectiveness of Visual Acoustic Fields in generating plausible impact sounds and accurately localizing impact sources. Our project page is at https://yuelei0428.github.io/projects/Visual-Acoustic-Fields/.

M3: 3D-Spatial MultiModal Memory

We present 3D Spatial MultiModal Memory (M3), a multimodal memory system designed to retain information about medium-sized static scenes through video sources for visual perception. By integrating 3D Gaussian Splatting techniques with foundation models, M3 builds a multimodal memory capable of rendering feature representations across granularities, encompassing a wide range of knowledge. In our exploration, we identify two key challenges in previous works on feature splatting: (1) computational constraints in storing high-dimensional features for each Gaussian primitive, and (2) misalignment or information loss between distilled features and foundation model features. To address these challenges, we propose M3 with key components of principal scene components and Gaussian memory attention, enabling efficient training and inference. To validate M3, we conduct comprehensive quantitative evaluations of feature similarity and downstream tasks, as well as qualitative visualizations to highlight the pixel trace of Gaussian memory attention. Our approach encompasses a diverse range of foundation models, including vision-language models (VLMs), perception models, and large multimodal and language models (LMMs/LLMs). Furthermore, to demonstrate real-world applicability, we deploy M3's feature field in indoor scenes on a quadruped robot. Notably, we claim that M3 is the first work to address the core compression challenges in 3D feature distillation.

Humanoid Policy ~ Human Policy

Training manipulation policies for humanoid robots with diverse data enhances their robustness and generalization across tasks and platforms. However, learning solely from robot demonstrations is labor-intensive, requiring expensive tele-operated data collection which is difficult to scale. This paper investigates a more scalable data source, egocentric human demonstrations, to serve as cross-embodiment training data for robot learning. We mitigate the embodiment gap between humanoids and humans from both the data and modeling perspectives. We collect an egocentric task-oriented dataset (PH2D) that is directly aligned with humanoid manipulation demonstrations. We then train a human-humanoid behavior policy, which we term Human Action Transformer (HAT). The state-action space of HAT is unified for both humans and humanoid robots and can be differentiably retargeted to robot actions. Co-trained with smaller-scale robot data, HAT directly models humanoid robots and humans as different embodiments without additional supervision. We show that human data improves both generalization and robustness of HAT with significantly better data collection efficiency. Code and data: https://human-as-robot.github.io/

WorldModelBench: Judging Video Generation Models As World Models

Video generation models have rapidly progressed, positioning themselves as video world models capable of supporting decision-making applications like robotics and autonomous driving. However, current benchmarks fail to rigorously evaluate these claims, focusing only on general video quality, ignoring important factors to world models such as physics adherence. To bridge this gap, we propose WorldModelBench, a benchmark designed to evaluate the world modeling capabilities of video generation models in application-driven domains. WorldModelBench offers two key advantages: (1) Against to nuanced world modeling violations: By incorporating instruction-following and physics-adherence dimensions, WorldModelBench detects subtle violations, such as irregular changes in object size that breach the mass conservation law - issues overlooked by prior benchmarks. (2) Aligned with large-scale human preferences: We crowd-source 67K human labels to accurately measure 14 frontier models. Using our high-quality human labels, we further fine-tune an accurate judger to automate the evaluation procedure, achieving 8.6% higher average accuracy in predicting world modeling violations than GPT-4o with 2B parameters. In addition, we demonstrate that training to align human annotations by maximizing the rewards from the judger noticeably improve the world modeling capability. The website is available at https://worldmodelbench-team.github.io.

RigAnything: Template-Free Autoregressive Rigging for Diverse 3D Assets

We present RigAnything, a novel autoregressive transformer-based model, which makes 3D assets rig-ready by probabilistically generating joints, skeleton topologies, and assigning skinning weights in a template-free manner. Unlike most existing auto-rigging methods, which rely on predefined skeleton template and are limited to specific categories like humanoid, RigAnything approaches the rigging problem in an autoregressive manner, iteratively predicting the next joint based on the global input shape and the previous prediction. While autoregressive models are typically used to generate sequential data, RigAnything extends their application to effectively learn and represent skeletons, which are inherently tree structures. To achieve this, we organize the joints in a breadth-first search (BFS) order, enabling the skeleton to be defined as a sequence of 3D locations and the parent index. Furthermore, our model improves the accuracy of position prediction by leveraging diffusion modeling, ensuring precise and consistent placement of joints within the hierarchy. This formulation allows the autoregressive model to efficiently capture both spatial and hierarchical relationships within the skeleton. Trained end-to-end on both RigNet and Objaverse datasets, RigAnything demonstrates state-of-the-art performance across diverse object types, including humanoids, quadrupeds, marine creatures, insects, and many more, surpassing prior methods in quality, robustness, generalizability, and efficiency. Please check our website for more details: https://www.liuisabella.com/RigAnything.

LoRA-TTT: Low-Rank Test-Time Training for Vision-Language Models

The rapid advancements in vision-language models (VLMs), such as CLIP, have intensified the need to address distribution shifts between training and testing datasets. Although prior Test-Time Training (TTT) techniques for VLMs have demonstrated robust performance, they predominantly rely on tuning text prompts, a process that demands substantial computational resources and is heavily dependent on entropy-based loss. In this paper, we propose LoRA-TTT, a novel TTT method that leverages Low-Rank Adaptation (LoRA), applied exclusively to the image encoder of VLMs. By introducing LoRA and updating only its parameters during test time, our method offers a simple yet effective TTT approach, retaining the model's initial generalization capability while achieving substantial performance gains with minimal memory and runtime overhead. Additionally, we introduce a highly efficient reconstruction loss tailored for TTT. Our method can adapt to diverse domains by combining these two losses, without increasing memory consumption or runtime. Extensive experiments on two benchmarks, covering 15 datasets, demonstrate that our method improves the zero-shot top-1 accuracy of CLIP-ViT-B/16 by an average of 5.79% on the OOD benchmark and 1.36% on the fine-grained benchmark, efficiently surpassing test-time prompt tuning, without relying on any external models or cache.

Integrating LMM Planners and 3D Skill Policies for Generalizable Manipulation

The recent advancements in visual reasoning capabilities of large multimodal models (LMMs) and the semantic enrichment of 3D feature fields have expanded the horizons of robotic capabilities. These developments hold significant potential for bridging the gap between high-level reasoning from LMMs and low-level control policies utilizing 3D feature fields. In this work, we introduce LMM-3DP, a framework that can integrate LMM planners and 3D skill Policies. Our approach consists of three key perspectives: high-level planning, low-level control, and effective integration. For high-level planning, LMM-3DP supports dynamic scene understanding for environment disturbances, a critic agent with self-feedback, history policy memorization, and reattempts after failures. For low-level control, LMM-3DP utilizes a semantic-aware 3D feature field for accurate manipulation. In aligning high-level and low-level control for robot actions, language embeddings representing the high-level policy are jointly attended with the 3D feature field in the 3D transformer for seamless integration. We extensively evaluate our approach across multiple skills and long-horizon tasks in a real-world kitchen environment. Our results show a significant 1.45x success rate increase in low-level control and an approximate 1.5x improvement in high-level planning accuracy compared to LLM-based baselines. Demo videos and an overview of LMM-3DP are available at https://lmm-3dp-release.github.io.

Diffusion Autoencoders are Scalable Image Tokenizers

Tokenizing images into compact visual representations is a key step in learning efficient and high-quality image generative models. We present a simple diffusion tokenizer (DiTo) that learns compact visual representations for image generation models. Our key insight is that a single learning objective, diffusion L2 loss, can be used for training scalable image tokenizers. Since diffusion is already widely used for image generation, our insight greatly simplifies training such tokenizers. In contrast, current state-of-the-art tokenizers rely on an empirically found combination of heuristics and losses, thus requiring a complex training recipe that relies on non-trivially balancing different losses and pretrained supervised models. We show design decisions, along with theoretical grounding, that enable us to scale DiTo for learning competitive image representations. Our results show that DiTo is a simpler, scalable, and self-supervised alternative to the current state-of-the-art image tokenizer which is supervised. DiTo achieves competitive or better quality than state-of-the-art in image reconstruction and downstream image generation tasks.

Parallel Sequence Modeling via Generalized Spatial Propagation Network

We present the Generalized Spatial Propagation Network (GSPN), a new attention mechanism optimized for vision tasks that inherently captures 2D spatial structures. Existing attention models, including transformers, linear attention, and state-space models like Mamba, process multi-dimensional data as 1D sequences, compromising spatial coherence and efficiency. GSPN overcomes these limitations by directly operating on spatially coherent image data and forming dense pairwise connections through a line-scan approach. Central to GSPN is the Stability-Context Condition, which ensures stable, context-aware propagation across 2D sequences and reduces the effective sequence length to $\sqrt{N}$ for a square map with N elements, significantly enhancing computational efficiency. With learnable, input-dependent weights and no reliance on positional embeddings, GSPN achieves superior spatial fidelity and state-of-the-art performance in vision tasks, including ImageNet classification, class-guided image generation, and text-to-image generation. Notably, GSPN accelerates SD-XL with softmax-attention by over $84\times$ when generating 16K images.