Songwei Ge

Niket Agarwal, Arslan Ali, Maciej Bala et al. · nvidia

Physical AI needs to be trained digitally first. It needs a digital twin of itself, the policy model, and a digital twin of the world, the world model. In this paper, we present the Cosmos World Foundation Model Platform to help developers build customized world models for their Physical AI setups. We position a world foundation model as a general-purpose world model that can be fine-tuned into customized world models for downstream applications. Our platform covers a video curation pipeline, pre-trained world foundation models, examples of post-training of pre-trained world foundation models, and video tokenizers. To help Physical AI builders solve the most critical problems of our society, we make Cosmos open-source and our models open-weight with permissive licenses available via https://github.com/nvidia-cosmos/cosmos-predict1.

Thomas Hayes, Songyang Zhang, Xi Yin et al.

Multimodal video-audio-text understanding and generation can benefit from datasets that are narrow but rich. The narrowness allows bite-sized challenges that the research community can make progress on. The richness ensures we are making progress along the core challenges. To this end, we present a large-scale video-audio-text dataset MUGEN, collected using the open-sourced platform game CoinRun [11]. We made substantial modifications to make the game richer by introducing audio and enabling new interactions. We trained RL agents with different objectives to navigate the game and interact with 13 objects and characters. This allows us to automatically extract a large collection of diverse videos and associated audio. We sample 375K video clips (3.2s each) and collect text descriptions from human annotators. Each video has additional annotations that are extracted automatically from the game engine, such as accurate semantic maps for each frame and templated textual descriptions. Altogether, MUGEN can help progress research in many tasks in multimodal understanding and generation. We benchmark representative approaches on tasks involving video-audio-text retrieval and generation. Our dataset and code are released at: https://mugen-org.github.io/.

Songwei Ge, Shlok Mishra, Simon Kornblith et al.

Although self-/un-supervised methods have led to rapid progress in visual representation learning, these methods generally treat objects and scenes using the same lens. In this paper, we focus on learning representations for objects and scenes that preserve the structure among them. Motivated by the observation that visually similar objects are close in the representation space, we argue that the scenes and objects should instead follow a hierarchical structure based on their compositionality. To exploit such a structure, we propose a contrastive learning framework where a Euclidean loss is used to learn object representations and a hyperbolic loss is used to encourage representations of scenes to lie close to representations of their constituent objects in a hyperbolic space. This novel hyperbolic objective encourages the scene-object hypernymy among the representations by optimizing the magnitude of their norms. We show that when pretraining on the COCO and OpenImages datasets, the hyperbolic loss improves downstream performance of several baselines across multiple datasets and tasks, including image classification, object detection, and semantic segmentation. We also show that the properties of the learned representations allow us to solve various vision tasks that involve the interaction between scenes and objects in a zero-shot fashion. Our code can be found at \url{https://github.com/shlokk/HCL/tree/main/HCL}.

Songwei Ge, Thomas Hayes, Harry Yang et al.

Videos are created to express emotion, exchange information, and share experiences. Video synthesis has intrigued researchers for a long time. Despite the rapid progress driven by advances in visual synthesis, most existing studies focus on improving the frames' quality and the transitions between them, while little progress has been made in generating longer videos. In this paper, we present a method that builds on 3D-VQGAN and transformers to generate videos with thousands of frames. Our evaluation shows that our model trained on 16-frame video clips from standard benchmarks such as UCF-101, Sky Time-lapse, and Taichi-HD datasets can generate diverse, coherent, and high-quality long videos. We also showcase conditional extensions of our approach for generating meaningful long videos by incorporating temporal information with text and audio. Videos and code can be found at https://songweige.github.io/projects/tats/index.html.

Songwei Ge, Taesung Park, Jun-Yan Zhu et al.

Plain text has become a prevalent interface for text-to-image synthesis. However, its limited customization options hinder users from accurately describing desired outputs. For example, plain text makes it hard to specify continuous quantities, such as the precise RGB color value or importance of each word. Furthermore, creating detailed text prompts for complex scenes is tedious for humans to write and challenging for text encoders to interpret. To address these challenges, we propose using a rich-text editor supporting formats such as font style, size, color, and footnote. We extract each word's attributes from rich text to enable local style control, explicit token reweighting, precise color rendering, and detailed region synthesis. We achieve these capabilities through a region-based diffusion process. We first obtain each word's region based on attention maps of a diffusion process using plain text. For each region, we enforce its text attributes by creating region-specific detailed prompts and applying region-specific guidance, and maintain its fidelity against plain-text generation through region-based injections. We present various examples of image generation from rich text and demonstrate that our method outperforms strong baselines with quantitative evaluations.

Quynh Phung, Songwei Ge, Jia-Bin Huang

Driven by the scalable diffusion models trained on large-scale datasets, text-to-image synthesis methods have shown compelling results. However, these models still fail to precisely follow the text prompt involving multiple objects, attributes, or spatial compositions. In this paper, we reveal the potential causes in the diffusion model's cross-attention and self-attention layers. We propose two novel losses to refocus attention maps according to a given spatial layout during sampling. Creating the layouts manually requires additional effort and can be tedious. Therefore, we explore using large language models (LLM) to produce these layouts for our method. We conduct extensive experiments on the DrawBench, HRS, and TIFA benchmarks to evaluate our proposed method. We show that our proposed attention refocusing effectively improves the controllability of existing approaches.

Ting-Hsuan Liao, Songwei Ge, Yiran Xu et al.

There has been tremendous progress in large-scale text-to-image synthesis driven by diffusion models enabling versatile downstream applications such as 3D object synthesis from texts, image editing, and customized generation. We present a generic approach using latent diffusion models as powerful image priors for various visual synthesis tasks. Existing methods that utilize such priors fail to use these models' full capabilities. To improve this, our core ideas are 1) a feature matching loss between features from different layers of the decoder to provide detailed guidance and 2) a KL divergence loss to regularize the predicted latent features and stabilize the training. We demonstrate the efficacy of our approach on three different applications, text-to-3D, StyleGAN adaptation, and layered image editing. Extensive results show our method compares favorably against baselines.

Songwei Ge, Shlok Mishra, Haohan Wang et al.

Unsupervised learning has recently made exceptional progress because of the development of more effective contrastive learning methods. However, CNNs are prone to depend on low-level features that humans deem non-semantic. This dependency has been conjectured to induce a lack of robustness to image perturbations or domain shift. In this paper, we show that by generating carefully designed negative samples, contrastive learning can learn more robust representations with less dependence on such features. Contrastive learning utilizes positive pairs that preserve semantic information while perturbing superficial features in the training images. Similarly, we propose to generate negative samples in a reversed way, where only the superfluous instead of the semantic features are preserved. We develop two methods, texture-based and patch-based augmentations, to generate negative samples. These samples achieve better generalization, especially under out-of-domain settings. We also analyze our method and the generated texture-based samples, showing that texture features are indispensable in classifying particular ImageNet classes and especially finer classes. We also show that model bias favors texture and shape features differently under different test settings. Our code, trained models, and ImageNet-Texture dataset can be found at https://github.com/SongweiGe/Contrastive-Learning-with-Non-Semantic-Negatives.

Songwei Ge, Vedanuj Goswami, C. Lawrence Zitnick et al.

Sketching or doodling is a popular creative activity that people engage in. However, most existing work in automatic sketch understanding or generation has focused on sketches that are quite mundane. In this work, we introduce two datasets of creative sketches -- Creative Birds and Creative Creatures -- containing 10k sketches each along with part annotations. We propose DoodlerGAN -- a part-based Generative Adversarial Network (GAN) -- to generate unseen compositions of novel part appearances. Quantitative evaluations as well as human studies demonstrate that sketches generated by our approach are more creative and of higher quality than existing approaches. In fact, in Creative Birds, subjects prefer sketches generated by DoodlerGAN over those drawn by humans! Our code can be found at https://github.com/facebookresearch/DoodlerGAN and a demo can be found at http://doodlergan.cloudcv.org.

Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar et al.

Fréchet Video Distance (FVD), a prominent metric for evaluating video generation models, is known to conflict with human perception occasionally. In this paper, we aim to explore the extent of FVD's bias toward per-frame quality over temporal realism and identify its sources. We first quantify the FVD's sensitivity to the temporal axis by decoupling the frame and motion quality and find that the FVD increases only slightly with large temporal corruption. We then analyze the generated videos and show that via careful sampling from a large set of generated videos that do not contain motions, one can drastically decrease FVD without improving the temporal quality. Both studies suggest FVD's bias towards the quality of individual frames. We further observe that the bias can be attributed to the features extracted from a supervised video classifier trained on the content-biased dataset. We show that FVD with features extracted from the recent large-scale self-supervised video models is less biased toward image quality. Finally, we revisit a few real-world examples to validate our hypothesis.

David McAllister, Songwei Ge, Brent Yi et al.

Flow-based generative models, including diffusion models, excel at modeling continuous distributions in high-dimensional spaces. In this work, we introduce Flow Policy Optimization (FPO), a simple on-policy reinforcement learning algorithm that brings flow matching into the policy gradient framework. FPO casts policy optimization as maximizing an advantage-weighted ratio computed from the conditional flow matching loss, in a manner compatible with the popular PPO-clip framework. It sidesteps the need for exact likelihood computation while preserving the generative capabilities of flow-based models. Unlike prior approaches for diffusion-based reinforcement learning that bind training to a specific sampling method, FPO is agnostic to the choice of diffusion or flow integration at both training and inference time. We show that FPO can train diffusion-style policies from scratch in a variety of continuous control tasks. We find that flow-based models can capture multimodal action distributions and achieve higher performance than Gaussian policies, particularly in under-conditioned settings.

Yue Feng, Vaibhav Sanjay, Spencer Lutz et al.

Automatically generating multiview illusions is a compelling challenge, where a single piece of visual content offers distinct interpretations from different viewing perspectives. Traditional methods, such as shadow art and wire art, create interesting 3D illusions but are limited to simple visual outputs (i.e., figure-ground or line drawing), restricting their artistic expressiveness and practical versatility. Recent diffusion-based illusion generation methods can generate more intricate designs but are confined to 2D images. In this work, we present a simple yet effective approach for creating 3D multiview illusions based on user-provided text prompts or images. Our method leverages a pre-trained text-to-image diffusion model to optimize the textures and geometry of neural 3D representations through differentiable rendering. When viewed from multiple angles, this produces different interpretations. We develop several techniques to improve the quality of the generated 3D multiview illusions. We demonstrate the effectiveness of our approach through extensive experiments and showcase illusion generation with diverse 3D forms.

Andrew Z. Wang, Songwei Ge, Tero Karras et al.

Both text-to-image generation and large language models (LLMs) have made significant advancements. However, many text-to-image models still employ the somewhat outdated T5 and CLIP as their text encoders. In this work, we investigate the effectiveness of using modern decoder-only LLMs as text encoders for text-to-image diffusion models. We build a standardized training and evaluation pipeline that allows us to isolate and evaluate the effect of different text embeddings. We train a total of 27 text-to-image models with 12 different text encoders to analyze the critical aspects of LLMs that could impact text-to-image generation, including the approaches to extract embeddings, different LLMs variants, and model sizes. Our experiments reveal that the de facto way of using last-layer embeddings as conditioning leads to inferior performance. Instead, we explore embeddings from various layers and find that using layer-normalized averaging across all layers significantly improves alignment with complex prompts. Most LLMs with this conditioning outperform the baseline T5 model, showing enhanced performance in advanced visio-linguistic reasoning skills.

Ting-Hsuan Liao, Haowen Liu, Yiran Xu et al.

We present PAD3R, a method for reconstructing deformable 3D objects from casually captured, unposed monocular videos. Unlike existing approaches, PAD3R handles long video sequences featuring substantial object deformation, large-scale camera movement, and limited view coverage that typically challenge conventional systems. At its core, our approach trains a personalized, object-centric pose estimator, supervised by a pre-trained image-to-3D model. This guides the optimization of deformable 3D Gaussian representation. The optimization is further regularized by long-term 2D point tracking over the entire input video. By combining generative priors and differentiable rendering, PAD3R reconstructs high-fidelity, articulated 3D representations of objects in a category-agnostic way. Extensive qualitative and quantitative results show that PAD3R is robust and generalizes well across challenging scenarios, highlighting its potential for dynamic scene understanding and 3D content creation.

David McAllister, Songwei Ge, Jia-Bin Huang et al.

Score distillation sampling (SDS) has proven to be an important tool, enabling the use of large-scale diffusion priors for tasks operating in data-poor domains. Unfortunately, SDS has a number of characteristic artifacts that limit its usefulness in general-purpose applications. In this paper, we make progress toward understanding the behavior of SDS and its variants by viewing them as solving an optimal-cost transport path from a source distribution to a target distribution. Under this new interpretation, these methods seek to transport corrupted images (source) to the natural image distribution (target). We argue that current methods' characteristic artifacts are caused by (1) linear approximation of the optimal path and (2) poor estimates of the source distribution. We show that calibrating the text conditioning of the source distribution can produce high-quality generation and translation results with little extra overhead. Our method can be easily applied across many domains, matching or beating the performance of specialized methods. We demonstrate its utility in text-to-2D, text-based NeRF optimization, translating paintings to real images, optical illusion generation, and 3D sketch-to-real. We compare our method to existing approaches for score distillation sampling and show that it can produce high-frequency details with realistic colors.

Quynh Phung, Songwei Ge, Jia-Bin Huang

Despite the advances in text-to-image synthesis, particularly with diffusion models, generating visual instructions that require consistent representation and smooth state transitions of objects across sequential steps remains a formidable challenge. This paper introduces a simple, training-free framework to tackle the issues, capitalizing on the advancements in diffusion models and large language models (LLMs). Our approach systematically integrates text comprehension and image generation to ensure visual instructions are visually appealing and maintain consistency and accuracy throughout the instruction sequence. We validate the effectiveness by testing multi-step instructions and comparing the text alignment and consistency with several baselines. Our experiments show that our approach can visualize coherent and visually pleasing instructions

Songwei Ge, Seungjun Nah, Guilin Liu et al.

Despite tremendous progress in generating high-quality images using diffusion models, synthesizing a sequence of animated frames that are both photorealistic and temporally coherent is still in its infancy. While off-the-shelf billion-scale datasets for image generation are available, collecting similar video data of the same scale is still challenging. Also, training a video diffusion model is computationally much more expensive than its image counterpart. In this work, we explore finetuning a pretrained image diffusion model with video data as a practical solution for the video synthesis task. We find that naively extending the image noise prior to video noise prior in video diffusion leads to sub-optimal performance. Our carefully designed video noise prior leads to substantially better performance. Extensive experimental validation shows that our model, Preserve Your Own Correlation (PYoCo), attains SOTA zero-shot text-to-video results on the UCF-101 and MSR-VTT benchmarks. It also achieves SOTA video generation quality on the small-scale UCF-101 benchmark with a $10\times$ smaller model using significantly less computation than the prior art.

Songwei Ge, Devi Parikh

We ask the question: to what extent can recent large-scale language and image generation models blend visual concepts? Given an arbitrary object, we identify a relevant object and generate a single-sentence description of the blend of the two using a language model. We then generate a visual depiction of the blend using a text-based image generation model. Quantitative and qualitative evaluations demonstrate the superiority of language models over classical methods for conceptual blending, and of recent large-scale image generation models over prior models for the visual depiction.

Songwei Ge, Vasu Singla, Ronen Basri et al.

Shift invariance is a critical property of CNNs that improves performance on classification. However, we show that invariance to circular shifts can also lead to greater sensitivity to adversarial attacks. We first characterize the margin between classes when a shift-invariant linear classifier is used. We show that the margin can only depend on the DC component of the signals. Then, using results about infinitely wide networks, we show that in some simple cases, fully connected and shift-invariant neural networks produce linear decision boundaries. Using this, we prove that shift invariance in neural networks produces adversarial examples for the simple case of two classes, each consisting of a single image with a black or white dot on a gray background. This is more than a curiosity; we show empirically that with real datasets and realistic architectures, shift invariance reduces adversarial robustness. Finally, we describe initial experiments using synthetic data to probe the source of this connection.

Austin Dill, Songwei Ge, Eunsu Kang et al.

In order to generate novel 3D shapes with machine learning, one must allow for interpolation. The typical approach for incorporating this creative process is to interpolate in a learned latent space so as to avoid the problem of generating unrealistic instances by exploiting the model's learned structure. The process of the interpolation is supposed to form a semantically smooth morphing. While this approach is sound for synthesizing realistic media such as lifelike portraits or new designs for everyday objects, it subjectively fails to directly model the unexpected, unrealistic, or creative. In this work, we present a method for learning how to interpolate point clouds. By encoding prior knowledge about real-world objects, the intermediate forms are both realistic and unlike any existing forms. We show not only how this method can be used to generate "creative" point clouds, but how the method can also be leveraged to generate 3D models suitable for sculpture.

Austin Dill, Chun-Liang Li, Songwei Ge et al.

In this work, we explore the idea that effective generative models for point clouds under the autoencoding framework must acknowledge the relationship between a continuous surface, a discretized mesh, and a set of points sampled from the surface. This view motivates a generative model that works by progressively deforming a uniform sphere until it approximates the goal point cloud. We review the underlying concepts leading to this conclusion from computer graphics and topology in differential geometry, and model the generation process as deformation via deep neural network parameterization. Finally, we show that this view of the problem produces a model that can generate quality meshes efficiently.

Songwei Ge, Zhicheng Dou, Zhengbao Jiang et al.

Search results personalization has become an effective way to improve the quality of search engines. Previous studies extracted information such as past clicks, user topical interests, query click entropy and so on to tailor the original ranking. However, few studies have taken into account the sequential information underlying previous queries and sessions. Intuitively, the order of issued queries is important in inferring the real user interests. And more recent sessions should provide more reliable personal signals than older sessions. In addition, the previous search history and user behaviors should influence the personalization of the current query depending on their relatedness. To implement these intuitions, in this paper we employ a hierarchical recurrent neural network to exploit such sequential information and automatically generate user profile from historical data. We propose a query-aware attention model to generate a dynamic user profile based on the input query. Significant improvement is observed in the experiment with data from a commercial search engine when compared with several traditional personalization models. Our analysis reveals that the attention model is able to attribute higher weights to more related past sessions after fine training.

Songwei Ge, Curtis Xuan, Ruihua Song et al.

Sound effects play an essential role in producing high-quality radio stories but require enormous labor cost to add. In this paper, we address the problem of automatically adding sound effects to radio stories with a retrieval-based model. However, directly implementing a tag-based retrieval model leads to high false positives due to the ambiguity of story contents. To solve this problem, we introduce a retrieval-based framework hybridized with a semantic inference model which helps to achieve robust retrieval results. Our model relies on fine-designed features extracted from the context of candidate triggers. We collect two story dubbing datasets through crowdsourcing to analyze the setting of adding sound effects and to train and test our proposed methods. We further discuss the importance of each feature and introduce several heuristic rules for the trade-off between precision and recall. Together with the text-to-speech technology, our results reveal a promising automatic pipeline on producing high-quality radio stories.

Songwei Ge, Austin Dill, Eunsu Kang et al.

We explore the intersection of human and machine creativity by generating sculptural objects through machine learning. This research raises questions about both the technical details of automatic art generation and the interaction between AI and people, as both artists and the audience of art. We introduce two algorithms for generating 3D point clouds and then discuss their actualization as sculpture and incorporation into a holistic art installation. Specifically, the Amalgamated DeepDream (ADD) algorithm solves the sparsity problem caused by the naive DeepDream-inspired approach and generates creative and printable point clouds. The Partitioned DeepDream (PDD) algorithm further allows us to explore more diverse 3D object creation by combining point cloud clustering algorithms and ADD.

Haohan Wang, Songwei Ge, Eric P. Xing et al.

Despite their renowned predictive power on i.i.d. data, convolutional neural networks are known to rely more on high-frequency patterns that humans deem superficial than on low-frequency patterns that agree better with intuitions about what constitutes category membership. This paper proposes a method for training robust convolutional networks by penalizing the predictive power of the local representations learned by earlier layers. Intuitively, our networks are forced to discard predictive signals such as color and texture that can be gleaned from local receptive fields and to rely instead on the global structures of the image. Across a battery of synthetic and benchmark domain adaptation tasks, our method confers improved generalization out of the domain. Also, to evaluate cross-domain transfer, we introduce ImageNet-Sketch, a new dataset consisting of sketch-like images, that matches the ImageNet classification validation set in categories and scale.

Chun-Liang Li, Eunsu Kang, Songwei Ge et al.

Our team of artists and machine learning researchers designed a creative algorithm that can generate authentic sculptural artworks. These artworks do not mimic any given forms and cannot be easily categorized into the dataset categories. Our approach extends DeepDream from images to 3D point clouds. The proposed algorithm, Amalgamated DeepDream (ADD), leverages the properties of point clouds to create objects with better quality than the naive extension. ADD presents promise for the creativity of machines, the kind of creativity that pushes artists to explore novel methods or materials and to create new genres instead of creating variations of existing forms or styles within one genre. For example, from Realism to Abstract Expressionism, or to Minimalism. Lastly, we present the sculptures that are 3D printed based on the point clouds created by ADD.