Denoising Diffusion Probabilistic Models (DDPMs) have achieved remarkable success in various image generation tasks compared with Generative Adversarial Nets (GANs). Recent work on semantic synthesis mainly follows the \emph{de facto} GAN-based approaches, which may lead to unsatisfactory quality or diversity of generated images. In this paper, we propose a novel framework based DDPM for synthesis. Unlike previous conditional diffusion model directly feeds layout and noisy as input U-Net...

Diffusion models have shown remarkable success in visual synthesis, but also raised concerns about potential abuse for malicious purposes. In this paper, we seek to build a detector telling apart real images from diffusion-generated images. We find that existing detectors struggle detect generated by diffusion models, even if include specific model their training data. To address issue, propose novel image representation called DIffusion Reconstruction Error (DIRE), which measures the error...

Contrastive learning shows great potential in unpaired image-to-image translation, but sometimes the translated results are poor quality and contents not preserved consistently. In this paper, we uncover that negative examples play a critical role performance of contrastive for image translation. The previous methods randomly sampled from patches different positions source image, which effective to push positive close query examples. To address issue, present instance-wise hard Negative...

Existing face forgery detection models try to discriminate fake images by detecting only spatial artifacts (e.g., generative artifacts, blending) or mainly temporal flickering, discontinuity). They may experience significant performance degradation when facing out-domain artifacts. In this paper, we propose capture both and in one model for detection. A simple idea is leverage a spatiotemporal (3D ConvNet). However, find that it easily rely on type of artifact ignore the other. To address...

We present SinDiffusion, leveraging denoising diffusion models to capture internal distribution of patches from a single natural image. SinDiffusion significantly improves the quality and diversity generated samples compared with existing GAN-based approaches. It is based on two core designs. First, trained model at scale instead multiple progressive growing scales which serves as default setting in prior work. This avoids accumulation errors, cause characteristic artifacts results. Second,...

Affordance-aware human insertion is a controllable synthesis task aimed at seamlessly integrating person into scene while aligning pose with contextual affordance and preserving visual identity. Previous methods, typically reliant on general framework of inpainting that injects all conditional information single branch, often struggle the complexities real-world contexts nuanced attributes figures. To this end, we present novel DIS entangled dual-branch for A ffordance-aware task, termed as...

We present SinDiffusion, leveraging denoising diffusion models to capture internal distribution of patches from a single natural image. SinDiffusion significantly improves the quality and diversity generated samples compared with existing GAN-based approaches. It is based on two core designs. First, trained model at scale instead multiple progressive growing scales which serves as default setting in prior work. This avoids accumulation errors, cause characteristic artifacts results. Second,...

Diffusion models have shown remarkable success in visual synthesis, but also raised concerns about potential abuse for malicious purposes. In this paper, we seek to build a detector telling apart real images from diffusion-generated images. We find that existing detectors struggle detect generated by diffusion models, even if include specific model their training data. To address issue, propose novel image representation called DIffusion Reconstruction Error (DIRE), which measures the error...

Hand gesture-to-gesture translation is a significant and interesting problem, which serves as key role in many applications, such sign language production. This task involves fine-grained structure understanding of the mapping between source target gestures. Current works follow data-driven paradigm based on sparse 2D joint representation. However, given insufficient representation capability joints, this easily leads to blurry generation results with incorrect structure. In paper, we...

In this work, we are dedicated to a new task, i.e., hand-object interaction image generation, which aims conditionally generate the under given hand, object and their status. This task is challenging research-worthy in many potential application scenarios, such as AR/VR games online shopping, etc. To address problem, propose novel HOGAN framework, utilizes expressive model-aware representation leverages its inherent topology build unified surface space. space, explicitly consider complex...

In this work, we are dedicated to text-guided image generation and propose a novel framework, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">i.e</i> ., CLIP2GAN, by leveraging CLIP model StyleGAN. The key idea of our CLIP2GAN is bridge the output feature embedding space input latent StyleGAN, which realized introducing mapping network. training stage, encode an with map code, further used reconstruct image. way, network optimized in...

Abstract We present ABACUS-R, a method based on deep learning for designing amino acid sequences that autonomously fold into given target backbone. This predicts the sidechain type of central residue from its 3-D local environment by using an encoder-decoder network trained with multi-task strategy. The environmental features encoded include types but not conformations sidechains surrounding residues. eliminates needs reconstructing and optimizing structures, drastically simplifies sequence...

The memory cores are essential for a system-on-a-chip (SOC). To test the cores, in this paper we propose generalized embedded pattern generator any march algorithm. Without loss of functionality algorithm, also present systematic procedure with short time complexity to reduce hardware cost generator.

Generative adversarial networks have been widely used in image synthesis recent years and the quality of generated has greatly improved. However, flexibility to control decouple facial attributes (e.g., eyes, nose, mouth) is still limited. In this paper, we propose a novel approach, called ChildGAN, generate child's according images parents with heredity prior. The main idea disentangle latent space pre-trained generation model precisely face child clear semantics. We use distances between...

In this paper we present a systematic procedure to integrate multiple march algorithms into universal embedded test pattern generator the various kinds of memory cores in system-on-a-chip. With low hardware overhead, satisfied high fault coverage can be achieved by using proposed generator.

Recent studies have achieved remarkable success using deep generative models for the image animation of an arbitrary object.However, previous methods synthesize animated results in a frame-by-frame manner, which is prone to producing flickering and temporally inconsistent results. In this paper, we propose novel self-supervised framework leveraging temporal information animation. Our processes video clip directly instead processing each frame independently. To achieve coherence video, design...

Existing face forgery detection models try to discriminate fake images by detecting only spatial artifacts (e.g., generative artifacts, blending) or mainly temporal flickering, discontinuity). They may experience significant performance degradation when facing out-domain artifacts. In this paper, we propose capture both and in one model for detection. A simple idea is leverage a spatiotemporal (3D ConvNet). However, find that it easily rely on type of artifact ignore the other. To address...

This paper addresses a more reliable and fault-tolerant version of the standard relay placement problem (RPP) in design deployment wireless sensor networks. Given set sensors Euclidean plane, 2-connected (2CRPP) is to place minimum number relays such that each can communicate with at least one relay, all jointly form network. Since 2CRPP proven be NP-hard, this we proposed polynomial time approximation algorithm for mathematically proved its ratio bound by (4+ε), worst case.