Learnings from Scaling Visual Tokenizers for Reconstruction and Generation

Visual tokenization via auto-encoding empowers state-of-the-art image and video generative models by compressing pixels into a latent space. Although scaling Transformer-based generators has been central to recent advances, the tokenizer component itself is rarely scaled, leaving open questions about how auto-encoder design choices influence both its objective of reconstruction and downstream generative performance. Our work aims to conduct an exploration of scaling in auto-encoders to fill in this blank. To facilitate this exploration, we replace the typical convolutional backbone with an enhanced Vision Transformer architecture for Tokenization (ViTok). We train ViTok on large-scale image and video datasets far exceeding ImageNet-1K, removing data constraints on tokenizer scaling. We first study how scaling the auto-encoder bottleneck affects both reconstruction and generation -- and find that while it is highly correlated with reconstruction, its relationship with generation is more complex. We next explored the effect of separately scaling the auto-encoders' encoder and decoder on reconstruction and generation performance. Crucially, we find that scaling the encoder yields minimal gains for either reconstruction or generation, while scaling the decoder boosts reconstruction but the benefits for generation are mixed. Building on our exploration, we design ViTok as a lightweight auto-encoder that achieves competitive performance with state-of-the-art auto-encoders on ImageNet-1K and COCO reconstruction tasks (256p and 512p) while outperforming existing auto-encoders on 16-frame 128p video reconstruction for UCF-101, all with 2-5x fewer FLOPs. When integrated with Diffusion Transformers, ViTok demonstrates competitive performance on image generation for ImageNet-1K and sets new state-of-the-art benchmarks for class-conditional video generation on UCF-101.

DirectorLLM for Human-Centric Video Generation

In this paper, we introduce DirectorLLM, a novel video generation model that employs a large language model (LLM) to orchestrate human poses within videos. As foundational text-to-video models rapidly evolve, the demand for high-quality human motion and interaction grows. To address this need and enhance the authenticity of human motions, we extend the LLM from a text generator to a video director and human motion simulator. Utilizing open-source resources from Llama 3, we train the DirectorLLM to generate detailed instructional signals, such as human poses, to guide video generation. This approach offloads the simulation of human motion from the video generator to the LLM, effectively creating informative outlines for human-centric scenes. These signals are used as conditions by the video renderer, facilitating more realistic and prompt-following video generation. As an independent LLM module, it can be applied to different video renderers, including UNet and DiT, with minimal effort. Experiments on automatic evaluation benchmarks and human evaluations show that our model outperforms existing ones in generating videos with higher human motion fidelity, improved prompt faithfulness, and enhanced rendered subject naturalness.

LinGen: Towards High-Resolution Minute-Length Text-to-Video Generation with Linear Computational Complexity

Text-to-video generation enhances content creation but is highly computationally intensive: The computational cost of Diffusion Transformers (DiTs) scales quadratically in the number of pixels. This makes minute-length video generation extremely expensive, limiting most existing models to generating videos of only 10-20 seconds length. We propose a Linear-complexity text-to-video Generation (LinGen) framework whose cost scales linearly in the number of pixels. For the first time, LinGen enables high-resolution minute-length video generation on a single GPU without compromising quality. It replaces the computationally-dominant and quadratic-complexity block, self-attention, with a linear-complexity block called MATE, which consists of an MA-branch and a TE-branch. The MA-branch targets short-to-long-range correlations, combining a bidirectional Mamba2 block with our token rearrangement method, Rotary Major Scan, and our review tokens developed for long video generation. The TE-branch is a novel TEmporal Swin Attention block that focuses on temporal correlations between adjacent tokens and medium-range tokens. The MATE block addresses the adjacency preservation issue of Mamba and improves the consistency of generated videos significantly. Experimental results show that LinGen outperforms DiT (with a 75.6% win rate) in video quality with up to 15$\times$ (11.5$\times$) FLOPs (latency) reduction. Furthermore, both automatic metrics and human evaluation demonstrate our LinGen-4B yields comparable video quality to state-of-the-art models (with a 50.5%, 52.1%, 49.1% win rate with respect to Gen-3, LumaLabs, and Kling, respectively). This paves the way to hour-length movie generation and real-time interactive video generation. We provide 68s video generation results and more examples in our project website: https://lineargen.github.io/.

Movie Gen: A Cast of Media Foundation Models

We present Movie Gen, a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. We also show additional capabilities such as precise instruction-based video editing and generation of personalized videos based on a user's image. Our models set a new state-of-the-art on multiple tasks: text-to-video synthesis, video personalization, video editing, video-to-audio generation, and text-to-audio generation. Our largest video generation model is a 30B parameter transformer trained with a maximum context length of 73K video tokens, corresponding to a generated video of 16 seconds at 16 frames-per-second. We show multiple technical innovations and simplifications on the architecture, latent spaces, training objectives and recipes, data curation, evaluation protocols, parallelization techniques, and inference optimizations that allow us to reap the benefits of scaling pre-training data, model size, and training compute for training large scale media generation models. We hope this paper helps the research community to accelerate progress and innovation in media generation models. All videos from this paper are available at https://go.fb.me/MovieGenResearchVideos.

Pixel-Space Post-Training of Latent Diffusion Models

Latent diffusion models (LDMs) have made significant advancements in the field of image generation in recent years. One major advantage of LDMs is their ability to operate in a compressed latent space, allowing for more efficient training and deployment. However, despite these advantages, challenges with LDMs still remain. For example, it has been observed that LDMs often generate high-frequency details and complex compositions imperfectly. We hypothesize that one reason for these flaws is due to the fact that all pre- and post-training of LDMs are done in latent space, which is typically $8 \times 8$ lower spatial-resolution than the output images. To address this issue, we propose adding pixel-space supervision in the post-training process to better preserve high-frequency details. Experimentally, we show that adding a pixel-space objective significantly improves both supervised quality fine-tuning and preference-based post-training by a large margin on a state-of-the-art DiT transformer and U-Net diffusion models in both visual quality and visual flaw metrics, while maintaining the same text alignment quality.

PSLF: A PID Controller-incorporated Second-order Latent Factor Analysis Model for Recommender System

A second-order-based latent factor (SLF) analysis model demonstrates superior performance in graph representation learning, particularly for high-dimensional and incomplete (HDI) interaction data, by incorporating the curvature information of the loss landscape. However, its objective function is commonly bi-linear and non-convex, causing the SLF model to suffer from a low convergence rate. To address this issue, this paper proposes a PID controller-incorporated SLF (PSLF) model, leveraging two key strategies: a) refining learning error estimation by incorporating the PID controller principles, and b) acquiring second-order information insights through Hessian-vector products. Experimental results on multiple HDI datasets indicate that the proposed PSLF model outperforms four state-of-the-art latent factor models based on advanced optimizers regarding convergence rates and generalization performance.

Computation-friendly Graph Neural Network Design by Accumulating Knowledge on Large Language Models

Graph Neural Networks (GNNs), like other neural networks, have shown remarkable success but are hampered by the complexity of their architecture designs, which heavily depend on specific data and tasks. Traditionally, designing proper architectures involves trial and error, which requires intensive manual effort to optimize various components. To reduce human workload, researchers try to develop automated algorithms to design GNNs. However, both experts and automated algorithms suffer from two major issues in designing GNNs: 1) the substantial computational resources expended in repeatedly trying candidate GNN architectures until a feasible design is achieved, and 2) the intricate and prolonged processes required for humans or algorithms to accumulate knowledge of the interrelationship between graphs, GNNs, and performance. To further enhance the automation of GNN architecture design, we propose a computation-friendly way to empower Large Language Models (LLMs) with specialized knowledge in designing GNNs, thereby drastically shortening the computational overhead and development cycle of designing GNN architectures. Our framework begins by establishing a knowledge retrieval pipeline that comprehends the intercorrelations between graphs, GNNs, and performance. This pipeline converts past model design experiences into structured knowledge for LLM reference, allowing it to quickly suggest initial model proposals. Subsequently, we introduce a knowledge-driven search strategy that emulates the exploration-exploitation process of human experts, enabling quick refinement of initial proposals within a promising scope. Extensive experiments demonstrate that our framework can efficiently deliver promising (e.g., Top-5.77%) initial model proposals for unseen datasets within seconds and without any prior training and achieve outstanding search performance in a few iterations.

Class-aware and Augmentation-free Contrastive Learning from Label Proportion

Learning from Label Proportion (LLP) is a weakly supervised learning scenario in which training data is organized into predefined bags of instances, disclosing only the class label proportions per bag. This paradigm is essential for user modeling and personalization, where user privacy is paramount, offering insights into user preferences without revealing individual data. LLP faces a unique difficulty: the misalignment between bag-level supervision and the objective of instance-level prediction, primarily due to the inherent ambiguity in label proportion matching. Previous studies have demonstrated deep representation learning can generate auxiliary signals to promote the supervision level in the image domain. However, applying these techniques to tabular data presents significant challenges: 1) they rely heavily on label-invariant augmentation to establish multi-view, which is not feasible with the heterogeneous nature of tabular datasets, and 2) tabular datasets often lack sufficient semantics for perfect class distinction, making them prone to suboptimality caused by the inherent ambiguity of label proportion matching. To address these challenges, we propose an augmentation-free contrastive framework TabLLP-BDC that introduces class-aware supervision (explicitly aware of class differences) at the instance level. Our solution features a two-stage Bag Difference Contrastive (BDC) learning mechanism that establishes robust class-aware instance-level supervision by disassembling the nuance between bag label proportions, without relying on augmentations. Concurrently, our model presents a pioneering multi-task pretraining pipeline tailored for tabular-based LLP, capturing intrinsic tabular feature correlations in alignment with label proportion distribution. Extensive experiments demonstrate that TabLLP-BDC achieves state-of-the-art performance for LLP in the tabular domain.

An Analysis on Quantizing Diffusion Transformers

Diffusion Models (DMs) utilize an iterative denoising process to transform random noise into synthetic data. Initally proposed with a UNet structure, DMs excel at producing images that are virtually indistinguishable with or without conditioned text prompts. Later transformer-only structure is composed with DMs to achieve better performance. Though Latent Diffusion Models (LDMs) reduce the computational requirement by denoising in a latent space, it is extremely expensive to inference images for any operating devices due to the shear volume of parameters and feature sizes. Post Training Quantization (PTQ) offers an immediate remedy for a smaller storage size and more memory-efficient computation during inferencing. Prior works address PTQ of DMs on UNet structures have addressed the challenges in calibrating parameters for both activations and weights via moderate optimization. In this work, we pioneer an efficient PTQ on transformer-only structure without any optimization. By analysing challenges in quantizing activations and weights for diffusion transformers, we propose a single-step sampling calibration on activations and adapt group-wise quantization on weights for low-bit quantization. We demonstrate the efficiency and effectiveness of proposed methods with preliminary experiments on conditional image generation.

VEnvision3D: A Synthetic Perception Dataset for 3D Multi-Task Model Research

Developing a unified multi-task foundation model has become a critical challenge in computer vision research. In the current field of 3D computer vision, most datasets only focus on single task, which complicates the concurrent training requirements of various downstream tasks. In this paper, we introduce VEnvision3D, a large 3D synthetic perception dataset for multi-task learning, including depth completion, segmentation, upsampling, place recognition, and 3D reconstruction. Since the data for each task is collected in the same environmental domain, sub-tasks are inherently aligned in terms of the utilized data. Therefore, such a unique attribute can assist in exploring the potential for the multi-task model and even the foundation model without separate training methods. Meanwhile, capitalizing on the advantage of virtual environments being freely editable, we implement some novel settings such as simulating temporal changes in the environment and sampling point clouds on model surfaces. These characteristics enable us to present several new benchmarks. We also perform extensive studies on multi-task end-to-end models, revealing new observations, challenges, and opportunities for future research. Our dataset and code will be open-sourced upon acceptance.