StructuRUL

structurul

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses 3.2 LATENT STRUCTURAL DIFFUSION MODEL To incorporate the body skeletons for pose control, the simplest way is by feature residual (Mou et al., 2023) or input concatenation (Ju et al., 2023b). However, three problems remain: 1) The sparse keypoints only depict the coarse human structure, while the fine-grained geometry and foregroundbackground relationship are ignored. Besides, the naive DM training is merely supervised by RGB signals, which fails to capture the inherent structural information. 2) The image RGB and structure representations are spatially aligned but substantially different in latent space. How to jointly model them remains challenging. 3) In contrast to the colorful RGB images, the structure maps are mostly monotonous with similar values in local regions, which are hard to learn by DMs (Lin et al., 2023). Unified Model for Simultaneous Denoising. Our solution to the first problem is to simultaneously denoise the depth and surface-normal along with the synthesized RGB image. We choose them as additional learning targets due to two reasons: 1) Depth and normal can be easily annotated for large-scale dataset, which are also used in recent controllable T2I generation (Zhang & Agrawala, 2023). 2) As two commonly-used structural guidance, they complement the spatial relationship and geometry information, where the depth (Deng et al., 2022), normal (Wang et al., 2022), or both (Yu et al., 2022b) are proven beneficial in recent 3D studies. To this end, a naive method is to train three separate networks to denoise the RGB, depth, and normal individually. But the spatial alignment between them is hard to preserve. Therefore, we propose to capture the joint distribution in a unified model by simultaneous denoising, which can be trained with simplified objective (Ho et al., 2020): Furthermore, we find that the number of shared modules can trade-off between the spatial alignment and distribution learning: On the one hand, more shared layers guarantee the more similar features of final output, leading to the paired texture and structure corresponding to the same image. On the other hand, the RGB, depth, and normal can be treated as different views of the same image, where predicting them from the same feature resembles an image-to-image translation task in essence. Empirically, we find the optimal design to replicate the conv_in, first DownBlock, last UpBlock, and conv_out for each expert branch, where each branch’s skip-connections are maintained separately (as depicted in Fig. 2). This yields both the spatial alignment and joint capture of image texture and structure. Note that such design is not limited to three targets, but can generalize to arbitrary number of paired distributions by simply involving more branches with little computation overhead. 3.3 STRUCTURE-GUIDED REFINER This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses 3.2 LATENT STRUCTURAL DIFFUSION MODEL To incorporate the body skeletons for pose control, the simplest way is by feature residual (Mou et al., 2023) or input concatenation (Ju et al., 2023b). However, three problems remain: 1) The sparse keypoints only depict the coarse human structure, while the fine-grained geometry and foregroundbackground relationship are ignored. Besides, the naive DM training is merely supervised by RGB signals, which fails to capture the inherent structural information. 2) The image RGB and structure representations are spatially aligned but substantially different in latent space. How to jointly model them remains challenging. 3) In contrast to the colorful RGB images, the structure maps are mostly monotonous with similar values in local regions, which are hard to learn by DMs (Lin et al., 2023). Unified Model for Simultaneous Denoising. Our solution to the first problem is to simultaneously denoise the depth and surface-normal along with the synthesized RGB image. We choose them as additional learning targets due to two reasons: 1) Depth and normal can be easily annotated for large-scale dataset, which are also used in recent controllable T2I generation (Zhang & Agrawala, 2023). 2) As two commonly-used structural guidance, they complement the spatial relationship and geometry information, where the depth (Deng et al., 2022), normal (Wang et al., 2022), or both (Yu et al., 2022b) are proven beneficial in recent 3D studies. To this end, a naive method is to train three separate networks to denoise the RGB, depth, and normal individually. But the spatial alignment between them is hard to preserve. Therefore, we propose to capture the joint distribution in a unified model by simultaneous denoising, which can be trained with simplified objective (Ho et al., 2020): Unified Model for Simultaneous Denoising. Furthermore, we find that the number of shared modules can trade-off between the spatial alignment and distribution learning: On the one hand, more shared layers guarantee the more similar features of final output, leading to the paired texture and structure corresponding to the same image. On the other hand, the RGB, depth, and normal can be treated as different views of the same image, where predicting them from the same feature resembles an image-to-image translation task in essence. Empirically, we find the optimal design to replicate the conv_in, first DownBlock, last UpBlock, and conv_out for each expert branch, where each branch’s skip-connections are maintained separately (as depicted in Fig. 2). This yields both the spatial alignment and joint capture of image texture and structure. Note that such design is not limited to three targets, but can generalize to arbitrary number of paired distributions by simply involving more branches with little computation overhead. 3.3 STRUCTURE-GUIDED REFINER This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

How HyperHuman Solves Structural Challenges with a Latent Diffusion Model

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses ABSTRACT Despite significant advances in large-scale text-to-image models, achieving hyperrealistic human image generation remains a desirable yet unsolved task. Existing models like Stable Diffusion and DALL·E 2 tend to generate human images with incoherent parts or unnatural poses. To tackle these challenges, our key insight is that human image is inherently structural over multiple granularities, from the coarse-level body skeleton to the fine-grained spatial geometry. Therefore, capturing such correlations between the explicit appearance and latent structure in one model is essential to generate coherent and natural human images. To this end, we propose a unified framework, HyperHuman, that generates in-the-wild human images of high realism and diverse layouts. Specifically, 1) we first build a largescale human-centric dataset, named HumanVerse, which consists of 340M images with comprehensive annotations like human pose, depth, and surface-normal. 2) Next, we propose a Latent Structural Diffusion Model that simultaneously denoises the depth and surface-normal along with the synthesized RGB image. Our model enforces the joint learning of image appearance, spatial relationship, and geometry in a unified network, where each branch in the model complements to each other with both structural awareness and textural richness. 3) Finally, to further boost the visual quality, we propose a Structure-Guided Refiner to compose the predicted conditions for more detailed generation of higher resolution. Extensive experiments demonstrate that our framework yields the state-of-the-art performance, generating hyper-realistic human images under diverse scenarios 1 INTRODUCTION Generating hyper-realistic human images from user conditions, e.g., text and pose, is of great importance to various applications, such as image animation (Liu et al., 2019) and virtual try-on (Wang et al., 2018). To this end, many efforts explore the task of controllable human image generation. Early methods either resort to variational auto-encoders (VAEs) in a reconstruction manner (Ren et al., 2020), or improve the realism by generative adversarial networks (GANs) (Siarohin et al., 2019). Though some of them create high-quality images (Zhang et al., 2022; Jiang et al., 2022), the unstable training and limited model capacity confine them to small datasets of low diversity. Recent emergence of diffusion models (DMs) (Ho et al., 2020) has set a new paradigm for realistic synthesis and become the predominant architecture in Generative AI (Dhariwal & Nichol, 2021). Nevertheless, the exemplar text-to-image (T2I) models like Stable Diffusion (Rombach et al., 2022) and DALL·E 2 (Ramesh et al., 2022) still struggle to create human images with coherent anatomy, e.g., arms and legs, and natural poses. The main reason lies in that human is articulated with nonrigid deformations, requiring structural information that can hardly be depicted by text prompts. To enable structural control for image generation, recent works like ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023) introduce a learnable branch to modulate the pre-trained DMs, e.g., Stable Diffusion, in a plug-and-play manner. However, these approaches suffer from the feature discrepancy between the main and auxiliary branches, leading to inconsistency between the control signals (e.g., pose maps) and the generated images. To address the issue, HumanSD (Ju et al., 2023b) proposes to directly input body skeleton into the diffusion U-Net by channel-wise concatenation. However, it is confined to generating artistic style images of limited diversity. Besides, human images are synthesized only with pose control, while other structural information like depth maps and surface-normal maps are not considered. In a nutshell, previous studies either take a singular control signal as input condition, or treat different control signals separately as independent guidance, instead of modeling the multi-level correlations between human appearance and different types of structural information. Realistic human generation with coherent structure remains unsolved. In this paper, we propose a unified framework HyperHuman to generate in-the-wild human images of high realism and diverse layouts. The key insight is that human image is inherently structural over multiple granularities, from the coarse-level body skeleton to fine-grained spatial geometry. Therefore, capturing such correlations between the explicit appearance and latent structure in one model is essential to generate coherent and natural human images. Specifically, we first establish a large-scale human-centric dataset called HumanVerse that contains 340M in-the-wild human images of high quality and diversity. It has comprehensive annotations, such as the coarse-level body skeletons, the fine-grained depth and surface-normal maps, and the high-level image captions and attributes. Based on this, two modules are designed for hyper-realistic controllable human image generation. In Latent Structural Diffusion Model, we augment the pre-trained diffusion backbone to simultaneously denoise the RGB, depth, and normal. Appropriate network layers are chosen to be replicated as structural expert branches, so that the model can both handle input/output of different domains, and guarantee the spatial alignment among the denoised textures and structures. Thanks to such dedicated design, the image appearance, spatial relationship, and geometry are jointly modeled within a unified network, where each branch is complementary to each other with both structural awareness and textural richness. To generate monotonous depth and surface-normal that have similar values in local regions, we utilize an improved noise schedule to eliminate low-frequency information leakage. The same timestep is sampled for each branch to achieve better learning and feature fusion. With the spatially-aligned structure maps, in Structure-Guided Refiner, we compose the predicted conditions for detailed generation of high resolution. Moreover, we design a robust conditioning scheme to mitigate the effect of error accumulation in our two-stage generation pipeline. To summarize, our main contributions are three-fold: 1) We propose a novel HyperHuman framework for in-the-wild controllable human image generation of high realism. A large-scale humancentric dataset HumanVerse is curated with comprehensive annotations like human pose, depth, and surface normal. As one of the earliest attempts in human generation foundation model, we hope to benefit future research. 2) We propose the Latent Structural Diffusion Model to jointly capture the image appearance, spatial relationship, and geometry in a unified framework. The Structure-Guided Refiner is further devised to compose the predicted conditions for generation of better visual quality and higher resolution. 3) Extensive experiments demonstrate that our HyperHuman yields the state-of-the-art performance, generating hyper-realistic human images under diverse scenarios. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses ABSTRACT Despite significant advances in large-scale text-to-image models, achieving hyperrealistic human image generation remains a desirable yet unsolved task. Existing models like Stable Diffusion and DALL·E 2 tend to generate human images with incoherent parts or unnatural poses. To tackle these challenges, our key insight is that human image is inherently structural over multiple granularities, from the coarse-level body skeleton to the fine-grained spatial geometry. Therefore, capturing such correlations between the explicit appearance and latent structure in one model is essential to generate coherent and natural human images. To this end, we propose a unified framework, HyperHuman, that generates in-the-wild human images of high realism and diverse layouts. Specifically, 1) we first build a largescale human-centric dataset, named HumanVerse, which consists of 340M images with comprehensive annotations like human pose, depth, and surface-normal. 2) Next, we propose a Latent Structural Diffusion Model that simultaneously denoises the depth and surface-normal along with the synthesized RGB image. Our model enforces the joint learning of image appearance, spatial relationship, and geometry in a unified network, where each branch in the model complements to each other with both structural awareness and textural richness. 3) Finally, to further boost the visual quality, we propose a Structure-Guided Refiner to compose the predicted conditions for more detailed generation of higher resolution. Extensive experiments demonstrate that our framework yields the state-of-the-art performance, generating hyper-realistic human images under diverse scenarios 1 INTRODUCTION Generating hyper-realistic human images from user conditions, e.g., text and pose, is of great importance to various applications, such as image animation (Liu et al., 2019) and virtual try-on (Wang et al., 2018). To this end, many efforts explore the task of controllable human image generation. Early methods either resort to variational auto-encoders (VAEs) in a reconstruction manner (Ren et al., 2020), or improve the realism by generative adversarial networks (GANs) (Siarohin et al., 2019). Though some of them create high-quality images (Zhang et al., 2022; Jiang et al., 2022), the unstable training and limited model capacity confine them to small datasets of low diversity. Recent emergence of diffusion models (DMs) (Ho et al., 2020) has set a new paradigm for realistic synthesis and become the predominant architecture in Generative AI (Dhariwal & Nichol, 2021). Nevertheless, the exemplar text-to-image (T2I) models like Stable Diffusion (Rombach et al., 2022) and DALL·E 2 (Ramesh et al., 2022) still struggle to create human images with coherent anatomy, e.g., arms and legs, and natural poses. The main reason lies in that human is articulated with nonrigid deformations, requiring structural information that can hardly be depicted by text prompts. To enable structural control for image generation, recent works like ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023) introduce a learnable branch to modulate the pre-trained DMs, e.g., Stable Diffusion, in a plug-and-play manner. However, these approaches suffer from the feature discrepancy between the main and auxiliary branches, leading to inconsistency between the control signals (e.g., pose maps) and the generated images. To address the issue, HumanSD (Ju et al., 2023b) proposes to directly input body skeleton into the diffusion U-Net by channel-wise concatenation. However, it is confined to generating artistic style images of limited diversity. Besides, human images are synthesized only with pose control, while other structural information like depth maps and surface-normal maps are not considered. In a nutshell, previous studies either take a singular control signal as input condition, or treat different control signals separately as independent guidance, instead of modeling the multi-level correlations between human appearance and different types of structural information. Realistic human generation with coherent structure remains unsolved. In this paper, we propose a unified framework HyperHuman to generate in-the-wild human images of high realism and diverse layouts. The key insight is that human image is inherently structural over multiple granularities, from the coarse-level body skeleton to fine-grained spatial geometry. Therefore, capturing such correlations between the explicit appearance and latent structure in one model is essential to generate coherent and natural human images. Specifically, we first establish a large-scale human-centric dataset called HumanVerse that contains 340M in-the-wild human images of high quality and diversity. It has comprehensive annotations, such as the coarse-level body skeletons, the fine-grained depth and surface-normal maps, and the high-level image captions and attributes. Based on this, two modules are designed for hyper-realistic controllable human image generation. In Latent Structural Diffusion Model, we augment the pre-trained diffusion backbone to simultaneously denoise the RGB, depth, and normal. Appropriate network layers are chosen to be replicated as structural expert branches, so that the model can both handle input/output of different domains, and guarantee the spatial alignment among the denoised textures and structures. Thanks to such dedicated design, the image appearance, spatial relationship, and geometry are jointly modeled within a unified network, where each branch is complementary to each other with both structural awareness and textural richness. To generate monotonous depth and surface-normal that have similar values in local regions, we utilize an improved noise schedule to eliminate low-frequency information leakage. The same timestep is sampled for each branch to achieve better learning and feature fusion. With the spatially-aligned structure maps, in Structure-Guided Refiner, we compose the predicted conditions for detailed generation of high resolution. Moreover, we design a robust conditioning scheme to mitigate the effect of error accumulation in our two-stage generation pipeline. To summarize, our main contributions are three-fold: 1) We propose a novel HyperHuman framework for in-the-wild controllable human image generation of high realism. A large-scale humancentric dataset HumanVerse is curated with comprehensive annotations like human pose, depth, and surface normal. As one of the earliest attempts in human generation foundation model, we hope to benefit future research. 2) We propose the Latent Structural Diffusion Model to jointly capture the image appearance, spatial relationship, and geometry in a unified framework. The Structure-Guided Refiner is further devised to compose the predicted conditions for generation of better visual quality and higher resolution. 3) Extensive experiments demonstrate that our HyperHuman yields the state-of-the-art performance, generating hyper-realistic human images under diverse scenarios. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

HyperHuman: Hyper-Realistic Human Generation with Latent Structural Diffusion

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.4 MORE USER STUDY DETAILS The study involves 25 participants and annotates for a total of 8236 images in the zero-shot MSCOCO 2014 validation human subset. They take 2-3 days to complete all the user study task, with a final review to examine the validity of human preference. Specifically, we conduct side-by-side comparisons between our generated results and each baseline model’s results. The asking question is “Considering both the image aesthetics and text-image alignment, which image is better? Prompt: .” The labelers are unaware of which image corresponds to which baseline, i.e., the place of two compared images are shuffled to achieve fair comparison without bias. We note that all the labelers are well-trained for such text-to-image generation comparison tasks, who have passed the examination on a test set and have experience in this kind of comparisons for over 50 times. Below, we include the user study rating details for our method vs. baseline models. Each labeler can click on four options: a) The left image is better, in this case the corresponding model will get +1 grade. b) The right image is better. c) NSFW, which means the prompt/image contain NSFW contents, in this case both models will get 0 grade. d) Hard Case, where the labelers find it hard to tell which one’s image quality is better, in this case both models will get +0.5 grade. The detailed comparison statistics are shown in Table 8, where we report the grades of HyperHuman vs. baseline methods. It can be clearly seen that our proposed framework is superior than all the existing models, with better image quality, realism, aesthetics, and text-image alignment. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses A.4 MORE USER STUDY DETAILS The study involves 25 participants and annotates for a total of 8236 images in the zero-shot MSCOCO 2014 validation human subset. They take 2-3 days to complete all the user study task, with a final review to examine the validity of human preference. Specifically, we conduct side-by-side comparisons between our generated results and each baseline model’s results. The asking question is “Considering both the image aesthetics and text-image alignment, which image is better? Prompt: .” The labelers are unaware of which image corresponds to which baseline, i.e., the place of two compared images are shuffled to achieve fair comparison without bias. “Considering both the image aesthetics and text-image alignment, which image is better? Prompt: .” We note that all the labelers are well-trained for such text-to-image generation comparison tasks, who have passed the examination on a test set and have experience in this kind of comparisons for over 50 times. Below, we include the user study rating details for our method vs. baseline models. Each labeler can click on four options: a) The left image is better, in this case the corresponding model will get +1 grade. b) The right image is better. c) NSFW, which means the prompt/image contain NSFW contents, in this case both models will get 0 grade. d) Hard Case, where the labelers find it hard to tell which one’s image quality is better, in this case both models will get +0.5 grade. The detailed comparison statistics are shown in Table 8, where we report the grades of HyperHuman vs. baseline methods. It can be clearly seen that our proposed framework is superior than all the existing models, with better image quality, realism, aesthetics, and text-image alignment. a) b) c) HyperHuman This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

HyperHuman Tops Image Generation Models in User Study

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.2 MORE IMPLEMENTATION DETAILS We report implementation details like training hyper-parameters, and model architecture in Tab. 6. A.3 MORE ABLATION STUDY RESULTS We implement additional ablation study experiments on the second stage Structure-Guided Refiner. Note that due to the training resource limit and the resolution discrepancy between MS-COCO real images (512 × 512) and high-quality renderings (1024 × 1024), we conduct several toy ablation experiments in the lightweight 512 × 512 variant of our model: 1) w/o random dropout, where the all the input conditions are not dropout or masked out during the conditional training stage. 2) Only Text, where not any structural prediction is input to the model and we only use the text prompt as condition. 3) Condition on p, where we only use human pose skeleton p as input condition to the refiner network. 4) Condition on d that uses depth map d as input condition. 5) Condition on n that uses surface-normal n as input condition. And their combinations of 6) Condition on p, d; 7) Condition on p, n; 8) Condition on d, n, to verify the impact of each condition and the necessity of using such multi-level hierarchical structural guidance for fine-grained generation. The results are reported in Tab. 7. We can see that the random dropout conditioning scheme is crucial for more robust training with better image quality, especially in the two-stage generation pipeline. Besides, the structural map/guidance contains geometry and spatial relationship information, which are beneficial to image generation of higher quality. Another interesting phenomenon is that only conditioned on surface-normal n is better than conditioned on both the pose skeleton p and depth map d, which aligns with our intuition that surface-normal conveys rich structural information that mostly cover coarse-level skeleton and depth map, except for the keypoint location and foregroundbackground relationship. Overall, we can conclude from ablation results that: 1) Each condition (i.e., pose skeleton, depth map, and surface-normal) is important for higher-quality and more aligned generation, which validates the necessity of our first-stage Latent Structural Diffusion Model to jointly capture them. 2) The random dropout scheme for robust conditioning can essentially bridge the train-test error accumulation in two-stage pipeline, leading to better image results. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses A.2 MORE IMPLEMENTATION DETAILS We report implementation details like training hyper-parameters, and model architecture in Tab. 6. A.3 MORE ABLATION STUDY RESULTS We implement additional ablation study experiments on the second stage Structure-Guided Refiner. Note that due to the training resource limit and the resolution discrepancy between MS-COCO real images (512 × 512) and high-quality renderings (1024 × 1024), we conduct several toy ablation experiments in the lightweight 512 × 512 variant of our model: 1) w/o random dropout, where the all the input conditions are not dropout or masked out during the conditional training stage. 2) Only Text, where not any structural prediction is input to the model and we only use the text prompt as condition. 3) Condition on p, where we only use human pose skeleton p as input condition to the refiner network. 4) Condition on d that uses depth map d as input condition. 5) Condition on n that uses surface-normal n as input condition. And their combinations of 6) Condition on p, d; 7) Condition on p, n; 8) Condition on d, n, to verify the impact of each condition and the necessity of using such multi-level hierarchical structural guidance for fine-grained generation. The results are reported in Tab. 7. We can see that the random dropout conditioning scheme is crucial for more robust training with better image quality, especially in the two-stage generation pipeline. Besides, the structural map/guidance contains geometry and spatial relationship information, which are beneficial to image generation of higher quality. Another interesting phenomenon is that only conditioned on surface-normal n is better than conditioned on both the pose skeleton p and depth map d, which aligns with our intuition that surface-normal conveys rich structural information that mostly cover coarse-level skeleton and depth map, except for the keypoint location and foregroundbackground relationship. Overall, we can conclude from ablation results that: 1) Each condition (i.e., pose skeleton, depth map, and surface-normal) is important for higher-quality and more aligned generation, which validates the necessity of our first-stage Latent Structural Diffusion Model to jointly capture them. 2) The random dropout scheme for robust conditioning can essentially bridge the train-test error accumulation in two-stage pipeline, leading to better image results. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

How Pose, Depth, and Surface-Normal Impact HyperHuman’s Image Quality

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A APPENDIX In this supplemental document, we provide more details of the following contents: 1) Additional quantitative results (Sec. A.1). 2) More implementation details like network architecture, hyperparameters, and training setups, etc (Sec. A.2). 3) More ablation study results (Sec. A.3). 4) More user study details (Sec. A.4). 5) The impact of random seed to our model to show the robustness of our method (Sec. A.5). 6) Boarder impact and the ethical consideration of this work (Sec. A.6). 7) More visual comparison results with recent T2I models (Sec. A.7). 8) More qualitative results of our model (Sec. A.8). 9) The asset licenses we use in this work (Sec. A.9). A.1 ADDITIONAL QUANTITATIVE RESULTS FID-CLIP Curves. Due to the page limit, we only show tiny-size FID-CLIP and FIDCLIP-CLIP curves in the main paper and omit the curves of HumanSD (Ju et al., 2023b) due to its too large FID and FIDCLIP results for reasonable axis scale. Here, we show a clearer version of FID-CLIP and FIDCLIP-CLIP curves in Fig. 4. As broadly proven in recent text-to-image studies (Rombach et al., 2022; Nichol et al., 2021; Saharia et al., 2022), the classifier-free guidance (CFG) plays an important role in trading-off image quality and diversity, where the CFG scales around 7.0 − 8.0 (corresponding to the bottom-right part of the curve) are the commonly-used choices in practice. We can see from Fig. 4 that our model can achieve a competitive CLIP Score while maintaining superior image quality results, showing the efficacy of our proposed HyperHuman framework. Human Preference-Related Metrics. As shown in recent text-to-image generation evaluation studies, conventional image quality metrics like FID (Heusel et al., 2017), KID (Binkowski et al. ´ , 2018) and text-image alignment CLIP Score (Radford et al., 2021) diverge a lot from the human preference (Kirstain et al., 2023). To this end, we adopt two very recent human preference-related metrics: 1) PickScore (Kirstain et al., 2023), which is trained on the side-by-side comparisons of two T2I models. 2) HPS (Human Preference Score) V2 (Wu et al., 2023), which takes the user like/dislike statistics for scoring model training. The evaluation results are reported in Tab. 4, which show that our framework performs better than the baselines. Although the improvement seems to be marginal, we find current human preference-related metrics to be highly biased: The scoring models are mostly trained on the synthetic data with highest resolution of 1024 × 1024, which makes them favor unrealistic images of 1024 resolution, as they rarely see real images of higher resolution in score model training. In spite of this, we still achieve superior quantitative and qualitative results on these two metrics and a comprehensive user study, outperforming all the baseline methods. Pose Accuracy Results on Different CFG Scales. We additionally report the pose accuracy results over different CFG scales. Specifically, we evaluate the conditional human generation methods of ControlNet (Zhang & Agrawala, 2023), T2I-Adapter (Mou et al., 2023), HumanSD (Ju et al., 2023b), and ours on four metrics Average Precision (AP), Average Recall (AR), clean AP (APclean), and clean AR (ARclean) as mentioned in Sec. 5.1. We report on CFG scales ranging from 4.0 to 13.0 in Tab. 5, where our method is constantly better in terms of pose accuracy and controllability. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses A APPENDIX In this supplemental document, we provide more details of the following contents: 1) Additional quantitative results (Sec. A.1). 2) More implementation details like network architecture, hyperparameters, and training setups, etc (Sec. A.2). 3) More ablation study results (Sec. A.3). 4) More user study details (Sec. A.4). 5) The impact of random seed to our model to show the robustness of our method (Sec. A.5). 6) Boarder impact and the ethical consideration of this work (Sec. A.6). 7) More visual comparison results with recent T2I models (Sec. A.7). 8) More qualitative results of our model (Sec. A.8). 9) The asset licenses we use in this work (Sec. A.9). A.1 ADDITIONAL QUANTITATIVE RESULTS FID-CLIP Curves. Due to the page limit, we only show tiny-size FID-CLIP and FIDCLIP-CLIP curves in the main paper and omit the curves of HumanSD (Ju et al., 2023b) due to its too large FID and FIDCLIP results for reasonable axis scale. Here, we show a clearer version of FID-CLIP and FIDCLIP-CLIP curves in Fig. 4. As broadly proven in recent text-to-image studies (Rombach et al., 2022; Nichol et al., 2021; Saharia et al., 2022), the classifier-free guidance (CFG) plays an important role in trading-off image quality and diversity, where the CFG scales around 7.0 − 8.0 (corresponding to the bottom-right part of the curve) are the commonly-used choices in practice. We can see from Fig. 4 that our model can achieve a competitive CLIP Score while maintaining superior image quality results, showing the efficacy of our proposed HyperHuman framework. FID-CLIP Curves. Human Preference-Related Metrics . As shown in recent text-to-image generation evaluation studies, conventional image quality metrics like FID (Heusel et al., 2017), KID (Binkowski et al. ´ , 2018) and text-image alignment CLIP Score (Radford et al., 2021) diverge a lot from the human preference (Kirstain et al., 2023). To this end, we adopt two very recent human preference-related metrics: 1) PickScore (Kirstain et al., 2023), which is trained on the side-by-side comparisons of two T2I models. 2) HPS (Human Preference Score) V2 (Wu et al., 2023), which takes the user like/dislike statistics for scoring model training. The evaluation results are reported in Tab. 4, which show that our framework performs better than the baselines. Although the improvement seems to be marginal, we find current human preference-related metrics to be highly biased: The scoring models are mostly trained on the synthetic data with highest resolution of 1024 × 1024, which makes them favor unrealistic images of 1024 resolution, as they rarely see real images of higher resolution in score model training. In spite of this, we still achieve superior quantitative and qualitative results on these two metrics and a comprehensive user study, outperforming all the baseline methods. Human Preference-Related Metrics Pose Accuracy Results on Different CFG Scales. We additionally report the pose accuracy results over different CFG scales. Specifically, we evaluate the conditional human generation methods of ControlNet (Zhang & Agrawala, 2023), T2I-Adapter (Mou et al., 2023), HumanSD (Ju et al., 2023b), and ours on four metrics Average Precision (AP), Average Recall (AR), clean AP (APclean), and clean AR (ARclean) as mentioned in Sec. 5.1. We report on CFG scales ranging from 4.0 to 13.0 in Tab. 5, where our method is constantly better in terms of pose accuracy and controllability. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

In-Depth Analysis of Human Image Generation Models

How HyperHuman Pushes the Boundaries of Realistic Human Image Generation

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses 5.2 ABLATION STUDY Simultaneous Denoise with Expert Branch. We explore whether latent structural diffusion model helps, and how many layers to replicate in the structural expert branches: 1) Denoise RGB, that only learns to denoise an image. 2) Denoise RGB + Depth, which also predicts depth. 3) Half DownBlock & UpBlock. We replicate half of the first DownBlock and the last UpBlock, which contains one down/up-sample ResBlock and one AttnBlock. 4) Two DownBlocks & UpBlocks, where we copy the first two DownBlocks and the last two UpBlocks. The results are shown in Tab. 2 (top), which prove that the joint learning of image appearance, spatial relationship, and geometry is beneficial. We also find that while fewer replicate layers give more spatially aligned results, the per-branch parameters are insufficient to capture distributions of each modality. In contrast, excessive replicate layers lead to less feature fusion across different targets, which fails to complement to each other branches. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses 5.2 ABLATION STUDY Simultaneous Denoise with Expert Branch. We explore whether latent structural diffusion model helps, and how many layers to replicate in the structural expert branches: 1) Denoise RGB, that only learns to denoise an image. 2) Denoise RGB + Depth, which also predicts depth. 3) Half DownBlock & UpBlock. We replicate half of the first DownBlock and the last UpBlock, which contains one down/up-sample ResBlock and one AttnBlock. 4) Two DownBlocks & UpBlocks, where we copy the first two DownBlocks and the last two UpBlocks. The results are shown in Tab. 2 (top), which prove that the joint learning of image appearance, spatial relationship, and geometry is beneficial. We also find that while fewer replicate layers give more spatially aligned results, the per-branch parameters are insufficient to capture distributions of each modality. In contrast, excessive replicate layers lead to less feature fusion across different targets, which fails to complement to each other branches. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

HyperHuman Ablation Study: Optimal Expert Branch Design for Improved Image Generation

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses 5.1 MAIN RESULTS Evaluation Metrics. We adopt commonly-used metrics to make comprehensive comparisons from three perspectives: 1) Image Quality. FID, KID, and FIDCLIP are used to reflect quality and diversity. 2) Text-Image Alignment, where the CLIP similarity between text and image embeddings is reported. 3) Pose Accuracy. We use the state-of-the-art pose estimator to extract poses from synthetic images and compare with the input (GT) pose conditions. The Average Precision (AP) and Average Recall (AR) are adopted to evaluate the pose alignment. Note that due to the noisy pose estimation of in-the-wild COCO, we also use APclean and ARclean to only evaluate on the three most salient persons. Quantitative Analysis. We report zero-shot evaluation results in Tab. 1. For all methods, we use the default CFG scale of 7.5, which well balances the quality and diversity with appealing results. Thanks to the structural awareness from expert branches, our proposed HyperHuman outperforms previous works by a clear margin, achieving the best results on image quality and pose accuracy metrics and ranks second on CLIP score. Note that SDXL (Podell et al., 2023) uses two text encoders with 3× larger UNet of more cross-attention layers, leading to superior text-image alignment. In spite of this, we still obtain an on-par CLIP score and surpass all the other baselines that have similar text encoder parameters. We also show the FID-CLIP and FIDCLIP-CLIP curves over multiple CFG scales in Fig. 3, where our model balances well between image quality and text-alignment, especially for the commonly-used CFG scales (bottom right). Please see Sec. A.1 for more quantitative results. Qualitative Analysis. Fig. 1 shows example results (top) and comparisons with baselines (bottom). We can generate both photo-realistic images and stylistic rendering, showing better realism, quality, diversity, and controllability. A comprehensive user study is further conducted as shown in Tab. 3, where the users prefer HyperHuman to the general and controllable T2I models. Please refer to Appendix A.4, A.7, and A.8 for more user study details, visual comparisons, and qualitative results. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses 5.1 MAIN RESULTS Evaluation Metrics. We adopt commonly-used metrics to make comprehensive comparisons from three perspectives: 1) Image Quality. FID, KID, and FIDCLIP are used to reflect quality and diversity. 2) Text-Image Alignment, where the CLIP similarity between text and image embeddings is reported. 3) Pose Accuracy. We use the state-of-the-art pose estimator to extract poses from synthetic images and compare with the input (GT) pose conditions. The Average Precision (AP) and Average Recall (AR) are adopted to evaluate the pose alignment. Note that due to the noisy pose estimation of in-the-wild COCO, we also use APclean and ARclean to only evaluate on the three most salient persons. Evaluation Metrics. Quantitative Analysis. We report zero-shot evaluation results in Tab. 1. For all methods, we use the default CFG scale of 7.5, which well balances the quality and diversity with appealing results. Thanks to the structural awareness from expert branches, our proposed HyperHuman outperforms previous works by a clear margin, achieving the best results on image quality and pose accuracy metrics and ranks second on CLIP score. Note that SDXL (Podell et al., 2023) uses two text encoders with 3× larger UNet of more cross-attention layers, leading to superior text-image alignment. In spite of this, we still obtain an on-par CLIP score and surpass all the other baselines that have similar text encoder parameters. We also show the FID-CLIP and FIDCLIP-CLIP curves over multiple CFG scales in Fig. 3, where our model balances well between image quality and text-alignment, especially for the commonly-used CFG scales (bottom right). Please see Sec. A.1 for more quantitative results. Quantitative Analysis. Qualitative Analysis. Fig. 1 shows example results (top) and comparisons with baselines (bottom). We can generate both photo-realistic images and stylistic rendering, showing better realism, quality, diversity, and controllability. A comprehensive user study is further conducted as shown in Tab. 3, where the users prefer HyperHuman to the general and controllable T2I models. Please refer to Appendix A.4, A.7, and A.8 for more user study details, visual comparisons, and qualitative results. Qualitative Analysis. HyperHuman This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

HyperHuman Sets New Benchmark for Realism and Pose Accuracy in Human Image Generation

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses 5 EXPERIMENTS Experimental Settings. For the comprehensive evaluation, we divide our comparisons into two settings: 1) Quantitative analysis. All the methods are tested on the same benchmark, using the same prompt with DDIM Scheduler (Song et al., 2020a) for 50 denoising steps to generate the same resolution images of 512 × 512. 2) Qualitative analysis. We generate high-resolution 1024 × 1024 results for each model with the officially provided best configurations, such as the prompt engineering, noise scheduler, and classifier-free guidance (CFG) scale. Note that we use the RGB output of the first-stage Latent Structural Diffusion Model for numerical comparison, while the improved results from the second-stage Structure-Guided Refiner are merely utilized for visual comparison. Datasets. We follow common practices in T2I generation (Yu et al., 2022a) and filter out a human subset from MS-COCO 2014 validation (Lin et al., 2014) for zero-shot evaluation. In particular, offthe-shelf human detector and pose estimator are used to obtain 8, 236 images with clearly-visible humans for evaluation. All the ground truth images are resized and center-cropped to 512 × 512. To guarantee fair comparisons, we train first-stage Latent Structural Diffusion on HumanVerse, which is a subset of public LAION-2B and COYO, to report quantitative metrics. In addition, an internal dataset is adopted to train second-stage Structure-Guided Refiner only for visually pleasing results. Comparison Methods. We compare with two categories of open-source SOTA works: 1) General T2I models, including SD (Rombach et al., 2022) (SD 1.x & 2.x), SDXL (Podell et al., 2023), and IF (DeepFloyd, 2023). 2) Controllable methods with pose condition. Notably, ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023) can handle multiple structural signals like canny, depth, and normal, where we take their skeleton-conditioned variant for comparison. HumanSD (Ju et al., 2023b) is the most recent work that specializes in pose-guided human generation. Implementation Details. We resize and random-crop the RGB, depth, and normal to the target resolution of each stage. To enforce the model with size and location awareness, the original image height/width and crop coordinates are embedded in a similar way to time embedding (Podell et al., 2023). Our code is developed based on diffusers (von Platen et al., 2022). 1) For the Latent Structural Diffusion, we fine-tune the whole UNet from the pretrained SD-2.0-base to v-prediction (Salimans & Ho, 2022) in 512 × 512 resolution. The DDIMScheduler with improved noise schedule is used for both training and sampling. We train on 128 80G NVIDIA A100 GPUs in a batch size of 2, 048 for one week. 2) For the Structure-Guided Refiner, we choose SDXL-1.0-base as the frozen backbone and fine-tune to ϵ-prediction for high-resolution synthesis of 1024 × 1024. We train on 256 80G NVIDIA A100 GPUs in a batch size of 2, 048 for one week. The overall framework is optimized with AdamW (Kingma & Ba, 2015) in 1e − 5 learning rate, and 0.01 weight decay. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses 5 EXPERIMENTS Experimental Settings. For the comprehensive evaluation, we divide our comparisons into two settings: 1) Quantitative analysis. All the methods are tested on the same benchmark, using the same prompt with DDIM Scheduler (Song et al., 2020a) for 50 denoising steps to generate the same resolution images of 512 × 512. 2) Qualitative analysis. We generate high-resolution 1024 × 1024 results for each model with the officially provided best configurations, such as the prompt engineering, noise scheduler, and classifier-free guidance (CFG) scale. Note that we use the RGB output of the first-stage Latent Structural Diffusion Model for numerical comparison, while the improved results from the second-stage Structure-Guided Refiner are merely utilized for visual comparison. Experimental Settings. Datasets. We follow common practices in T2I generation (Yu et al., 2022a) and filter out a human subset from MS-COCO 2014 validation (Lin et al., 2014) for zero-shot evaluation. In particular, offthe-shelf human detector and pose estimator are used to obtain 8, 236 images with clearly-visible humans for evaluation. All the ground truth images are resized and center-cropped to 512 × 512. To guarantee fair comparisons, we train first-stage Latent Structural Diffusion on HumanVerse, which is a subset of public LAION-2B and COYO, to report quantitative metrics. In addition, an internal dataset is adopted to train second-stage Structure-Guided Refiner only for visually pleasing results. Datasets. Comparison Methods. We compare with two categories of open-source SOTA works: 1) General T2I models, including SD (Rombach et al., 2022) (SD 1.x & 2.x), SDXL (Podell et al., 2023), and IF (DeepFloyd, 2023). 2) Controllable methods with pose condition. Notably, ControlNet (Zhang & Agrawala, 2023) and T2I-Adapter (Mou et al., 2023) can handle multiple structural signals like canny, depth, and normal, where we take their skeleton-conditioned variant for comparison. HumanSD (Ju et al., 2023b) is the most recent work that specializes in pose-guided human generation. Comparison Methods. Implementation Details. We resize and random-crop the RGB, depth, and normal to the target resolution of each stage. To enforce the model with size and location awareness, the original image height/width and crop coordinates are embedded in a similar way to time embedding (Podell et al., 2023). Our code is developed based on diffusers (von Platen et al., 2022). 1) For the Latent Structural Diffusion, we fine-tune the whole UNet from the pretrained SD-2.0-base to v-prediction (Salimans & Ho, 2022) in 512 × 512 resolution. The DDIMScheduler with improved noise schedule is used for both training and sampling. We train on 128 80G NVIDIA A100 GPUs in a batch size of 2, 048 for one week. 2) For the Structure-Guided Refiner, we choose SDXL-1.0-base as the frozen backbone and fine-tune to ϵ-prediction for high-resolution synthesis of 1024 × 1024. We train on 256 80G NVIDIA A100 GPUs in a batch size of 2, 048 for one week. The overall framework is optimized with AdamW (Kingma & Ba, 2015) in 1e − 5 learning rate, and 0.01 weight decay. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

HyperHuman Outperforms ControlNet, SDXL, and Others in Groundbreaking Image Generation Tests

Authors:
(1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.;
(2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com;
(3) Aliaksandr Siarohin, Snap Inc.;
(4) Ivan Skorokhodov, Snap Inc.;
(5) Yanyu Li, Snap Inc.;
(6) Dahua Lin, CUHK;
(7) Xihui Liu, HKU;
(8) Ziwei Liu, NTU;
(9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 4 Human Verse Dataset 5 Experiments 5.1 Main Results 5.2 Ablation Study 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses 4 HUMANVERSE DATASET Large-scale datasets with high quality samples, rich annotations, and diverse distribution are crucial for image generation tasks (Schuhmann et al., 2022; Podell et al., 2023), especially in the human domain (Liu et al., 2016;Jiang et al., 2023). To facilitate controllable human generation of high-fidelity, we establish a comprehensive human dataset with extensive annotations named HumanVerse. Please kindly refer to Appendix A.9 for more details about the dataset and annotation resources we use. Dataset Preprocessing. We curate from two principled datasets: LAION-2B-en (Schuhmann et al., 2022) and COYO-700M (Byeon et al., 2022). To isolate human images, we employ YOLOS (Fang et al., 2021) for human detection. Specifically, only those images containing 1 to 3 human bounding boxes are retained, where people should be visible with an area ratio exceeding 15%. We further rule out samples of poor aesthetics (< 4.5) or low resolution (< 200 × 200). This yields a high-quality subset by eliminating blurry and over-small humans. Unlike existing models that mostly train on full-body humans of simple context (Zhang & Agrawala, 2023), our dataset encompasses a wider spectrum, including various backgrounds and partial human regions such as clothing and limbs. 2D Human Poses. 2D human poses (skeleton of joint keypoints), which serve as one of the most flexible and easiest obtainable coarse-level condition signals, are widely used in controllable human generation studies (Ju et al., 2023b). To achieve accurate keypoint annotations, we resort to MMPose (Contributors, 2020) as inference interface and choose ViTPose-H (Xu et al., 2022) as backbone that performs best over several pose estimation benchmarks. In particular, the per-instance bounding box, keypoint coordinates and confidence are labeled, including whole-body skeleton (133 keypoints), body skeleton (17 keypoints), hand (21 keypoints), and facial landmarks (68 keypoints). Depth and Surface-Normal Maps are fine-grained structures that reflect the spatial geometry of images (Wu et al., 2022), which are commonly used in conditional generation (Mou et al., 2023). We apply Omnidata (Eftekhar et al., 2021) for monocular depth and normal. The MiDaS (Ranftl et al., 2022) is further annotated following recent depth-to-image pipelines (Rombach et al., 2022). Outpaint for Accurate Annotations. Diffusion models have shown promising results on image inpainting and outpainting, where the appearance and structure of unseen regions can be hallucinated based on the visible parts. Motivated by this, we propose to outpaint each image for a more holistic view given that most off-the-shelf structure estimators are trained on the “complete” image views. Although the outpainted region may be imperfect with artifacts, it can complement a more comprehensive human structure. To this end, we utilize the powerful SD-Inpaint to outpaint the surrounding areas of the original canvas. These images are further processed by off-the-shelf estimators, where we only use the labeling within the original image region for more accurate annotations. Overall Statistics. In summary, COYO subset contains 90, 948, 474 (91M) images and LAION-2B subset contains 248, 396, 109 (248M) images, which is 18.12% and 20.77% of fullset, respectively. The whole annotation process takes 640 16/32G NVIDIA V100 GPUs for two weeks in parallel. This paper is available on arxiv under CC BY 4.0 DEED license. Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Authors: Authors: (1) Xian Liu, Snap Inc., CUHK with Work done during an internship at Snap Inc.; (2) Jian Ren, Snap Inc. with Corresponding author: jren@snapchat.com; (3) Aliaksandr Siarohin, Snap Inc.; (4) Ivan Skorokhodov, Snap Inc.; (5) Yanyu Li, Snap Inc.; (6) Dahua Lin, CUHK; (7) Xihui Liu, HKU; (8) Ziwei Liu, NTU; (9) Sergey Tulyakov, Snap Inc. Table of Links Abstract and 1 Introduction Abstract and 1 Introduction 2 Related Work 2 Related Work 3 Our Approach and 3.1 Preliminaries and Problem Setting 3 Our Approach and 3.1 Preliminaries and Problem Setting 3.2 Latent Structural Diffusion Model 3.2 Latent Structural Diffusion Model 3.3 Structure-Guided Refiner 3.3 Structure-Guided Refiner 4 Human Verse Dataset 4 Human Verse Dataset 5 Experiments 5 Experiments 5.1 Main Results 5.1 Main Results 5.2 Ablation Study 5.2 Ablation Study 6 Discussion and References 6 Discussion and References A Appendix and A.1 Additional Quantitative Results A Appendix and A.1 Additional Quantitative Results A.2 More Implementation Details and A.3 More Ablation Study Results A.2 More Implementation Details and A.3 More Ablation Study Results A.4 More User Study Details A.4 More User Study Details A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.5 Impact of Random Seed and Model Robustness and A.6 Boarder Impact and Ethical Consideration A.7 More Comparison Results and A.8 Additional Qualitative Results A.7 More Comparison Results and A.8 Additional Qualitative Results A.9 Licenses A.9 Licenses 4 HUMANVERSE DATASET Large-scale datasets with high quality samples, rich annotations, and diverse distribution are crucial for image generation tasks (Schuhmann et al., 2022; Podell et al., 2023), especially in the human domain (Liu et al., 2016;Jiang et al., 2023). To facilitate controllable human generation of high-fidelity, we establish a comprehensive human dataset with extensive annotations named HumanVerse. Please kindly refer to Appendix A.9 for more details about the dataset and annotation resources we use. Dataset Preprocessing. We curate from two principled datasets: LAION-2B-en (Schuhmann et al., 2022) and COYO-700M (Byeon et al., 2022). To isolate human images, we employ YOLOS (Fang et al., 2021) for human detection. Specifically, only those images containing 1 to 3 human bounding boxes are retained, where people should be visible with an area ratio exceeding 15%. We further rule out samples of poor aesthetics (< 4.5) or low resolution (< 200 × 200). This yields a high-quality subset by eliminating blurry and over-small humans. Unlike existing models that mostly train on full-body humans of simple context (Zhang & Agrawala, 2023), our dataset encompasses a wider spectrum, including various backgrounds and partial human regions such as clothing and limbs. Dataset Preprocessing. 2D Human Poses. 2D human poses (skeleton of joint keypoints), which serve as one of the most flexible and easiest obtainable coarse-level condition signals, are widely used in controllable human generation studies (Ju et al., 2023b). To achieve accurate keypoint annotations, we resort to MMPose (Contributors, 2020) as inference interface and choose ViTPose-H (Xu et al., 2022) as backbone that performs best over several pose estimation benchmarks. In particular, the per-instance bounding box, keypoint coordinates and confidence are labeled, including whole-body skeleton (133 keypoints), body skeleton (17 keypoints), hand (21 keypoints), and facial landmarks (68 keypoints). 2D Human Poses. Depth and Surface-Normal Maps are fine-grained structures that reflect the spatial geometry of images (Wu et al., 2022), which are commonly used in conditional generation (Mou et al., 2023). We apply Omnidata (Eftekhar et al., 2021) for monocular depth and normal. The MiDaS (Ranftl et al., 2022) is further annotated following recent depth-to-image pipelines (Rombach et al., 2022). Depth and Surface-Normal Maps Outpaint for Accurate Annotations. Diffusion models have shown promising results on image inpainting and outpainting, where the appearance and structure of unseen regions can be hallucinated based on the visible parts. Motivated by this, we propose to outpaint each image for a more holistic view given that most off-the-shelf structure estimators are trained on the “complete” image views. Although the outpainted region may be imperfect with artifacts, it can complement a more comprehensive human structure. To this end, we utilize the powerful SD-Inpaint to outpaint the surrounding areas of the original canvas. These images are further processed by off-the-shelf estimators, where we only use the labeling within the original image region for more accurate annotations. Outpaint for Accurate Annotations. Overall Statistics. In summary, COYO subset contains 90, 948, 474 (91M) images and LAION-2B subset contains 248, 396, 109 (248M) images, which is 18.12% and 20.77% of fullset, respectively. The whole annotation process takes 640 16/32G NVIDIA V100 GPUs for two weeks in parallel. This paper is available on arxiv under CC BY 4.0 DEED license. This paper is available on arxiv under CC BY 4.0 DEED license. available on arxiv

Stories, News & Mentions about STRUCTURUL

Structurul’s Stories on HackerNoon