TinyGAN: Distilling BigGAN for Conditional Image Generation

Chang, Ting-Yun; Lu, Chi-Jen

doi:10.1007/978-3-030-69538-5_31

Cited by 9 publications

(6 citation statements)

References 8 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…BigGAN 18 is a type of Generative Adversarial Network (GAN) developed by researchers at DeepMind in 2018. It is an extension of the StyleGAN architecture and is trained on a much larger dataset of images, allowing it to generate high‐resolution images with much more detail and realism.…”

Section: Related Workmentioning

confidence: 99%

Diversified realistic face image generation GAN for human subjects in multimedia content creation

Kumar,

Singh

2024

Computer Animation & Virtual

View full text Add to dashboard Cite

Face image generation plays an important role in generating innovative and unique multimedia content using the GAN model. With these qualities of the GAN model, they have numerous challenges in the human face image generation. The problems encountered in the generation of facial images are like blurriness in images, incomplete details in the generated facial images, high computational power requirements, and so forth. In this manuscript, we proposed a GAN model that utilizes the composite strength of VGG‐16 and ResNet‐50's models to overcome those difficulties. It uses VGG‐16 to build a discriminator model to discriminate between real and fake images. The generator model utilizes a combination of components from the ResNet‐50 and VGG‐16 models to enhance the image generation process at each iteration, resulting in the creation of realistic face images. The proposed DRFI GAN (Diversified and Realistic Face Image Generation GAN) model's generator achieves an impressive low FID score of 20.50, which is less than existing state‐of‐the‐art approaches. Furthermore, our findings indicate that the images generated by the DRFI GAN model exhibit 10%–15% greater efficiency and realism with reduced training time compared to existing state‐of‐the‐art methods with lower FID scores.

show abstract

Section: Related Workmentioning

confidence: 99%

Diversified realistic face image generation GAN for human subjects in multimedia content creation

Kumar,

Singh

2024

Computer Animation & Virtual

View full text Add to dashboard Cite

show abstract

“…SAGAN 11 (Self‐Attention GAN) incorporates self‐attention mechanisms to improve image synthesis quality. BigGAN 12 is a GAN model that achieves state‐of‐the‐art results on image generation tasks by scaling up the size and capacity of the generator network.…”

Section: Related Workmentioning

confidence: 99%

Contour wavelet diffusion: A fast and high‐quality image generation model

Ding,

Zhu,

Zou

2024

Computational Intelligence

View full text Add to dashboard Cite

Diffusion models can generate high‐quality images and have attracted increasing attention. However, diffusion models adopt a progressive optimization process and often have long training and inference time, which limits their application in realistic scenarios. Recently, some latent space diffusion models have partially accelerated training speed by using parameters in the feature space, but additional network structures still require a large amount of unnecessary computation. Therefore, we propose the Contour Wavelet Diffusion method to accelerate the training and inference speed. First, we introduce the contour wavelet transform to extract anisotropic low‐frequency and high‐frequency components from the input image, and achieve acceleration by processing these down‐sampling components. Meanwhile, due to the good reconstructive properties of wavelet transforms, the quality of generated images can be maintained. Second, we propose a Batch‐normalized stochastic attention module that enables the model to effectively focus on important high‐frequency information, further improving the quality of image generation. Finally, we propose a balanced loss function to further improve the convergence speed of the model. Experimental results on several public datasets show that our method can significantly accelerate the training and inference speed of the diffusion model while ensuring the quality of generated images.

show abstract

“…GAN Compression. We highlight a few recent methods among many GAN compression methods [28,5,6,32,22,10]. GAN Slimming [32] integrates model distillation, channel pruning and quantization into a unified framework.…”

Section: Related Workmentioning

confidence: 99%

“…For style mixing, we replace the i−th vector in w A with that from w B . We set i ∈ [1,3], i ∈ [5,8] and i ∈ [10,13] for coarse, middle and fine style mixing, respectively. For interpolation, we linearly combine the latent code with β controls the weight: w = β • w A + (1 − β) • w B , and then feed w into generator to get the interpolation results.…”

Section: A4 Image Editingmentioning

confidence: 99%

Mind the Gap in Distilling StyleGANs

Xu¹,

Hou²,

Liu³

et al. 2022

Preprint

View full text Add to dashboard Cite

StyleGAN family is one of the most popular Generative Adversarial Networks (GANs) for unconditional generation. Despite its impressive performance, its high demand on storage and computation impedes their deployment on resource-constrained devices. This paper provides a comprehensive study of distilling from the popular StyleGANlike architecture. Our key insight is that the main challenge of Style-GAN distillation lies in the output discrepancy issue, where the teacher and student model yield different outputs given the same input latent code. Standard knowledge distillation losses typically fail under this heterogeneous distillation scenario. We conduct thorough analysis about the reasons and effects of this discrepancy issue, and identify that the mapping network plays a vital role in determining semantic information of generated images. Based on this finding, we propose a novel initialization strategy for the student model, which can ensure the output consistency to the maximum extent. To further enhance the semantic consistency between the teacher and student model, we present a latentdirection-based distillation loss that preserves the semantic relations in latent space. Extensive experiments demonstrate the effectiveness of our approach in distilling StyleGAN2 and StyleGAN3, outperforming existing GAN distillation methods by a large margin. Code is available at: https://github.com/xuguodong03/StyleKD

show abstract

TinyGAN: Distilling BigGAN for Conditional Image Generation

Cited by 9 publications

References 8 publications

Diversified realistic face image generation GAN for human subjects in multimedia content creation

Diversified realistic face image generation GAN for human subjects in multimedia content creation

Contour wavelet diffusion: A fast and high‐quality image generation model

Mind the Gap in Distilling StyleGANs

Contact Info

Product

Resources

About