NeurIPS 2024 Papers

Dichotomous Image Segmentation (DIS) tasks require highly precise annotations, and traditional dataset creation methods are labor intensive, costly, and require extensive domain expertise. Although using synthetic data for DIS is a promising solution to these challenges, current generative models and techniques struggle with the issues of scene deviations, noise-induced errors, and limited training sample variability. To address these issues, we introduce a novel approach, Mask Factory, which provides a scalable solution for generating diverse and precise datasets, markedly reducing preparation time and costs. We first introduce a general mask editing method that combines rigid and non-rigid editing techniques to generate high-quality synthetic masks. Specially, rigid editing leverages geometric priors from diffusion models to achieve precise viewpoint transformations under zero-shot conditions, while non-rigid editing employs adversarial training and self-attention mechanisms for complex, topologically consistent modifications. Then, we generate pairs of high-resolution image and accurate segmentation mask using a multi-conditional control generation method. Finally, our experiments on the widely-used DIS5K dataset benchmark demonstrate superior performance in quality and efficiency compared to existing methods. The code is available at https://qian-hao-tian.github.io/MaskFactory/.

The rapid advancements in the development of multimodal large language models (MLLMs) have consistently led to new breakthroughs on various benchmarks. In response, numerous challenging and comprehensive benchmarks have been proposed to more accurately assess the capabilities of MLLMs. However, there is a dearth of exploration of the higher-order perceptual capabilities of MLLMs. To fill this gap, we propose the Image Implication understanding Benchmark, II-Bench, which aims to evaluate the model's higher-order perception of images. Through extensive experiments on II-Bench across multiple MLLMs, we have made significant findings. Initially, a substantial gap is observed between the performance of MLLMs and humans on II-Bench. The pinnacle accuracy of MLLMs attains 74.8%, whereas human accuracy averages 90%, peaking at an impressive 98%. Subsequently, MLLMs perform worse on abstract and complex images, suggesting limitations in their ability to understand high-level semantics and capture image details. Finally, it is observed that most models exhibit enhanced accuracy when image sentiment polarity hints are incorporated into the prompts. This observation underscores a notable deficiency in their inherent understanding of image sentiment. We believe that II-Bench will inspire the community to develop the next generation of MLLMs, advancing the journey towards expert artificial general intelligence (AGI). II-Bench is publicly available at https://huggingface.co/datasets/m-a-p/II-Bench.

The task of layout-to-image generation involves synthesizing images based on the captions of objects and their spatial positions. Existing methods still struggle in complex layout generation, where common bad cases include object missing, inconsistent lighting, conflicting view angles, etc. To effectively address these issues, we propose a \textbf{Hi}erarchical \textbf{Co}ntrollable (HiCo) diffusion model for layout-to-image generation, featuring object seperable conditioning branch structure. Our key insight is to achieve spatial disentanglement through hierarchical modeling of layouts. We use a multi branch structure to represent hierarchy and aggregate them in fusion module. To evaluate the performance of multi-objective controllable layout generation in natural scenes, we introduce the HiCo-7K benchmark, derived from the GRIT-20M dataset and manually cleaned. https://github.com/360CVGroup/HiCo_T2I.

Rectified flow and reflow procedures have significantly advanced fast generation by progressively straightening ordinary differential equation (ODE) flows under the assumption that image and noise pairs, known as coupling, can be approximated by straight trajectories with constant velocity. However, we observe that the constant velocity modeling and reflow procedures have limitations in accurately learning to couple with flow crossing, leading to suboptimal few-step generation. To overcome the limitations, we introduce the Constant Acceleration Flow (CAF), a novel framework based on a simple constant acceleration equation. Additionally, we propose two techniques to improve estimation accuracy: initial velocity conditioning for the acceleration model and a reflow process for the initial velocity. Our comparative studies show that CAF not only outperforms rectified flow with reflow procedures in terms of speed and accuracy but also demonstrates substantial improvements in preserving coupling for fast generation.

Human image animation aims to generate a human motion video from the inputs of a reference human image and a target motion video. Current diffusion-based image animation systems exhibit high precision in transferring human identity into targeted motion, yet they still exhibit irregular quality in their outputs. Their optimal precision is achieved only when the physical compositions (i.e., scale and rotation) of the human shapes in the reference image and target pose frame are aligned. In the absence of such alignment, there is a noticeable decline in fidelity and consistency. Especially, in real-world environments, this compositional misalignment commonly occurs, posing significant challenges to the practical usage of current systems. To this end, we propose Test-time Procrustes Calibration (TPC), which enhances the robustness of diffusion-based image animation systems by maintaining optimal performance even when faced with compositional misalignment, effectively addressing real-world scenarios. The TPC provides a calibrated reference image for the diffusion model, enhancing its capability to understand the correspondence between human shapes in the reference and target images. Our method is simple and can be applied to any diffusion-based image animation system in a model-agnostic manner, improving the effectiveness at test time without additional training.

The diffusion model performs remarkable in generating high-dimensional content but is computationally intensive, especially during training. We propose Progressive Growing of Diffusion Autoencoder (PaGoDA), a novel pipeline that reduces the training costs through three stages: training diffusion on downsampled data, distilling the pretrained diffusion, and progressive super-resolution. With the proposed pipeline, PaGoDA achieves a $64\times$ reduced cost in training its diffusion model on $8\times$ downsampled data; while at the inference, with the single-step, it performs state-of-the-art on ImageNet across all resolutions from $64\times64$ to $512\times512$, and text-to-image. PaGoDA's pipeline can be applied directly in the latent space, adding compression alongside the pre-trained autoencoder in Latent Diffusion Models (e.g., Stable Diffusion). The code is available at https://github.com/sony/pagoda.

Diffusion models face significant challenges when employed for large-scale medical image reconstruction in real practice such as 3D Computed Tomography (CT). Due to the demanding memory, time, and data requirements, it is difficult to train a diffusion model directly on the entire volume of high-dimensional data to obtain an efficient 3D diffusion prior. Existing works utilizing diffusion priors on single 2D image slice with hand-crafted cross-slice regularization would sacrifice the z-axis consistency, which results in severe artifacts along the z-axis. In this work, we propose a novel framework that enables learning the 3D image prior through position-aware 3D-patch diffusion score blending for reconstructing large-scale 3D medical images. To the best of our knowledge, we are the first to utilize a 3D-patch diffusion prior for 3D medical image reconstruction. Extensive experiments on sparse view and limited angle CT reconstruction show that our DiffusionBlend method significantly outperforms previous methods and achieves state-of-the-art performance on real-world CT reconstruction problems with high-dimensional 3D image (i.e., $256 \times 256 \times 500$). Our algorithm also comes with better or comparable computational efficiency than previous state-of-the-art methods. Code is available at https://github.com/efzero/DiffusionBlend.

We introduce a doubly hierarchical generative representation for strand-based 3D hairstyle geometry that progresses from coarse, low-pass filtered guide hair to densely populated hair strands rich in high-frequency details. We employ the Discrete Cosine Transform (DCT) to separate low-frequency structural curves from high-frequency curliness and noise, avoiding the Gibbs' oscillation issues associated with the standard Fourier transform in open curves. Unlike the guide hair sampled from the scalp UV map grids which may lose capturing details of the hairstyle in existing methods, our method samples optimal sparse guide strands by utilising $k$-medoids clustering centres from low-pass filtered dense strands, which more accurately retain the hairstyle's inherent characteristics. The proposed variational autoencoder-based generation network, with an architecture inspired by geometric deep learning and implicit neural representations, facilitates flexible, off-the-grid guide strand modelling and enables the completion of dense strands in any quantity and density, drawing on principles from implicit neural representations. Empirical evaluations confirm the capacity of the model to generate convincing guide hair and dense strands, complete with nuanced high-frequency details.

We propose a novel method, \textbf{TwinAct}, to tackle the challenge of decoupling actions and actors in order to customize the text-guided diffusion models (TGDMs) for few-shot action image generation. TwinAct addresses the limitations of existing methods that struggle to decouple actions from other semantics (e.g., the actor's appearance) due to the lack of an effective inductive bias with few exemplar images. Our approach introduces a common action space, which is a textual embedding space focused solely on actions, enabling precise customization without actor-related details. Specifically, TwinAct involves three key steps: 1) Building common action space based on a set of representative action phrases; 2) Imitating the customized action within the action space; and 3) Generating highly adaptable customized action images in diverse contexts with action similarity loss. To comprehensively evaluate TwinAct, we construct a novel benchmark, which provides sample images with various forms of actions. Extensive experiments demonstrate TwinAct's superiority in generating accurate, context-independent customized actions while maintaining the identity consistency of different subjects, including animals, humans, and even customized actors.

Designing a fast and effective entropy model is challenging but essential for practical application of neural codecs. Beyond spatial autoregressive entropy models, more efficient backward adaptation-based entropy models have been recently developed. They not only reduce decoding time by using smaller number of modeling steps but also maintain or even improve rate--distortion performance by leveraging more diverse contexts for backward adaptation. Despite their significant progress, we argue that their performance has been limited by the simple adoption of the design convention for forward adaptation: using only a single type of hyper latent representation, which does not provide sufficient contextual information, especially in the first modeling step. In this paper, we propose a simple yet effective entropy modeling framework that leverages sufficient contexts for forward adaptation without compromising on bit-rate. Specifically, we introduce a strategy of diversifying hyper latent representations for forward adaptation, i.e., using two additional types of contexts along with the existing single type of context. In addition, we present a method to effectively use the diverse contexts for contextualizing the current elements to be encoded/decoded. By addressing the limitation of the previous approach, our proposed framework leads to significant performance improvements. Experimental results on popular datasets show that our proposed framework consistently improves rate-distortion performance across various bit-rate regions, e.g., 3.73\% BD-rate gain over the state-of-the-art baseline on the Kodak dataset.

We present Visual AutoRegressive modeling (VAR), a new generation paradigm that redefines the autoregressive learning on images as coarse-to-fine "next-scale prediction" or "next-resolution prediction", diverging from the standard raster-scan "next-token prediction". This simple, intuitive methodology allows autoregressive (AR) transformers to learn visual distributions fast and generalize well: VAR, for the first time, makes GPT-style AR models surpass diffusion transformers in image generation. On ImageNet 256x256 benchmark, VAR significantly improve AR baseline by improving Frechet inception distance (FID) from 18.65 to 1.73, inception score (IS) from 80.4 to 350.2, with around 20x faster inference speed. It is also empirically verified that VAR outperforms the Diffusion Transformer (DiT) in multiple dimensions including image quality, inference speed, data efficiency, and scalability. Scaling up VAR models exhibits clear power-law scaling laws similar to those observed in LLMs, with linear correlation coefficients near -0.998 as solid evidence. VAR further showcases zero-shot generalization ability in downstream tasks including image in-painting, out-painting, and editing. These results suggest VAR has initially emulated the two important properties of LLMs: Scaling Laws and zero-shot task generalization. We have released all models and codes to promote the exploration of AR/VAR models for visual generation and unified learning.

Trustworthiness of generative language models (GLMs) is crucial in their deployment to critical decision making systems. Hence, certified risk control methods such as selective prediction and conformal prediction have been applied to mitigating the hallucination problem in various supervised downstream tasks. However, the lack of appropriate correctness metric hinders applying such principled methods to language generation tasks. In this paper, we circumvent this problem by leveraging the concept of textual entailment to evaluate the correctness of the generated sequence, and propose two selective generation algorithms which control the false discovery rate with respect to the textual entailment relation (FDR-E) with a theoretical guarantee: $\texttt{SGen}^{\texttt{Sup}}$ and $\texttt{SGen}^{\texttt{Semi}}$. $\texttt{SGen}^{\texttt{Sup}}$, a direct modification of the selective prediction, is a supervised learning algorithm which exploits entailment-labeled data, annotated by humans. Since human annotation is costly, we further propose a semi-supervised version, $\texttt{SGen}^{\texttt{Semi}}$, which fully utilizes the unlabeled data by pseudo-labeling, leveraging an entailment set function learned via conformal prediction. Furthermore, $\texttt{SGen}^{\texttt{Semi}}$ enables to use more general class of selection functions, neuro-selection functions, and provides users with an optimal selection function class given multiple candidates. Finally, we demonstrate the efficacy of the $\texttt{SGen}$ family in achieving a desired FDR-E level with comparable selection efficiency to those from baselines on both open and closed source GLMs. Code and datasets are provided at https://github.com/ml-postech/selective-generation.

Recent advancements in models linking natural language with human motions have shown significant promise in motion generation and editing based on instructional text. Motivated by applications in sports coaching and motor skill learning, we investigate the inverse problem: generating corrective instructional text, leveraging motion editing and generation models. We introduce a novel approach that, given a user’s current motion (source) and the desired motion (target), generates text instructions to guide the user towards achieving the target motion. We leverage large language models to generate corrective texts and utilize existing motion generation and editing frameworks to compile datasets of triplets (source motion, target motion, and corrective text). Using this data, we propose a new motion-language model for generating corrective instructions. We present both qualitative and quantitative results across a diverse range of applications that largely improve upon baselines. Our approach demonstrates its effectiveness in instructional scenarios, offering text-based guidance to correct and enhance user performance.

Diffusion models have demonstrated remarkable success in various domains, including molecular generation. However, conditional molecular generation remains a fundamental challenge due to an intrinsic trade-off between targeting specific chemical properties and generating meaningful samples from the data distribution. In this work, we present Time-Aware Conditional Synthesis (TACS), a novel approach to conditional generation on diffusion models. It integrates adaptively controlled plug-and-play "online" guidance into a diffusion model, driving samples toward the desired properties while maintaining validity and stability. A key component of our algorithm is our new type of diffusion sampler, Time Correction Sampler (TCS), which is used to control guidance and ensure that the generated molecules remain on the correct manifold at each reverse step of the diffusion process at the same time. Our proposed method demonstrates significant performance in conditional 3D molecular generation and offers a promising approach towards inverse molecular design, potentially facilitating advancements in drug discovery, materials science, and other related fields.

Different camera sensors have different noise patterns, and thus an image denoising model trained on one sensor often does not generalize well to a different sensor. One plausible solution is to collect a large dataset for each sensor for training or fine-tuning, which is inevitably time-consuming. To address this cross-domain challenge, we present a novel adaptive domain learning (ADL) scheme for cross-domain RAW image denoising by utilizing existing data from different sensors (source domain) plus a small amount of data from the new sensor (target domain). The ADL training scheme automatically removes the data in the source domain that are harmful to fine-tuning a model for the target domain (some data are harmful as adding them during training lowers the performance due to domain gaps). Also, we introduce a modulation module to adopt sensor-specific information (sensor type and ISO) to understand input data for image denoising. We conduct extensive experiments on public datasets with various smartphone and DSLR cameras, which show our proposed model outperforms prior work on cross-domain image denoising, given a small amount of image data from the target domain sensor.

Existing super-resolution (SR) methods optimize all model weights equally using $\mathcal{L}_1$ or $\mathcal{L}_2$ losses by uniformly sampling image patches without considering dataset imbalances or parameter redundancy, which limits their performance. To address this, we formulate the image SR task as an imbalanced distribution transfer learning problem from a statistical probability perspective, proposing a plug-and-play Weight-Balancing framework (WBSR) to achieve balanced model learning without changing the original model structure and training data. Specifically, we develop a Hierarchical Equalization Sampling (HES) strategy to address data distribution imbalances, enabling better feature representation from texture-rich samples. To tackle model optimization imbalances, we propose a Balanced Diversity Loss (BDLoss) function, focusing on learning texture regions while disregarding redundant computations in smooth regions. After joint training of HES and BDLoss to rectify these imbalances, we present a gradient projection dynamic inference strategy to facilitate accurate and efficient inference. Extensive experiments across various models, datasets, and scale factors demonstrate that our method achieves comparable or superior performance to existing approaches with about 34\% reduction in computational cost.

In this work, we present TextHarmony, a unified and versatile multimodal generative model proficient in comprehending and generating visual text. Simultaneously generating images and texts typically results in performance degradation due to the inherent inconsistency between vision and language modalities. To overcome this challenge, existing approaches resort to modality-specific data for supervised fine-tuning, necessitating distinct model instances. We propose Slide-LoRA, which dynamically aggregates modality-specific and modality-agnostic LoRA experts, partially decoupling the multimodal generation space. Slide-LoRA harmonizes the generation of vision and language within a singular model instance, thereby facilitating a more unified generative process. Additionally, we develop a high-quality image caption dataset, DetailedTextCaps-100K, synthesized with a sophisticated closed-source MLLM to enhance visual text generation capabilities further. Comprehensive experiments across various benchmarks demonstrate the effectiveness of the proposed approach. Empowered by Slide-LoRA, TextHarmony achieves comparable performance to modality-specific fine-tuning results with only a 2% increase in parameters and shows an average improvement of 2.5% in visual text comprehension tasks and 4.0% in visual text generation tasks. Our work delineates the viability of an integrated approach to multimodal generation within the visual text domain, setting a foundation for subsequent inquiries. Code is available at https://github.com/bytedance/TextHarmony.

Prediction is a fundamental capability of all living organisms, and has been proposed as an objective for learning sensory representations. Recent work demonstrates that in primate visual systems, prediction is facilitated by neural representations that follow straighter temporal trajectories than their initial photoreceptor encoding, which allows for prediction by linear extrapolation. Inspired by these experimental findings, we develop a self-supervised learning (SSL) objective that explicitly quantifies and promotes straightening. We demonstrate the power of this objective in training deep feedforward neural networks on smoothly-rendered synthetic image sequences that mimic commonly-occurring properties of natural videos. The learned model contains neural embeddings that are predictive, but also factorize the geometric, photometric, and semantic attributes of objects. The representations also prove more robust to noise and adversarial attacks compared to previous SSL methods that optimize for invariance to random augmentations. Moreover, these beneficial properties can be transferred to other training procedures by using the straightening objective as a regularizer, suggesting a broader utility for straightening as a principle for robust unsupervised learning.

3D generation methods have shown visually compelling results powered by diffusion image priors. However, they often fail to produce realistic geometric details, resulting in overly smooth surfaces or geometric details inaccurately baked in albedo maps. To address this, we introduce a new method that incorporates touch as an additional modality to improve the geometric details of generated 3D assets. We design a lightweight 3D texture field to synthesize visual and tactile textures, guided by 2D diffusion model priors on both visual and tactile domains. We condition the visual texture generation on high-resolution tactile normals and guide the patch-based tactile texture refinement with a customized TextureDreambooth. We further present a multi-part generation pipeline that enables us to synthesize different textures across various regions. To our knowledge, we are the first to leverage high-resolution tactile sensing to enhance geometric details for 3D generation tasks. We evaluate our method in both text-to-3D and image-to-3D settings. Our experiments demonstrate that our method provides customized and realistic fine geometric textures while maintaining accurate alignment between two modalities of vision and touch.