基于 PyTorch 和 MMCV的图像和视频生成工具箱，支持多种生成模型如图像生成，条件生成，图像转换+源代码+文档说明

# StyleGANv2 > [Analyzing and Improving the Image Quality of Stylegan](https://openaccess.thecvf.com/content_CVPR_2020/html/Karras_Analyzing_and_Improving_the_Image_Quality_of_StyleGAN_CVPR_2020_paper.html) <!-- [ALGORITHM] --> ## Abstract <!-- [ABSTRACT] --> The style-based GAN architecture (StyleGAN) yields state-of-the-art results in data-driven unconditional generative image modeling. We expose and analyze several of its characteristic artifacts, and propose changes in both model architecture and training methods to address them. In particular, we redesign the generator normalization, revisit progressive growing, and regularize the generator to encourage good conditioning in the mapping from latent codes to images. In addition to improving image quality, this path length regularizer yields the additional benefit that the generator becomes significantly easier to invert. This makes it possible to reliably attribute a generated image to a particular network. We furthermore visualize how well the generator utilizes its output resolution, and identify a capacity problem, motivating us to train larger models for additional quality improvements. Overall, our improved model redefines the state of the art in unconditional image modeling, both in terms of existing distribution quality metrics as well as perceived image quality. <!-- [IMAGE] --> <div align=center> <img src="https://user-images.githubusercontent.com/28132635/143055738-58b15493-ab87-436d-94fd-c4c19e4a0225.JPG"/> </div> ## Results and Models <div align="center"> <b> Results (compressed) from StyleGAN2 config-f trained by MMGeneration</b> <br/> <img src="https://user-images.githubusercontent.com/12726765/113825919-25433100-97b4-11eb-84f7-5c66b3cfbc68.png" width="800"/> </div> | Model | Comment | FID50k | P&R50k | Config | Download | | :---------------------------------: | :-------------: | :----: | :-----------: | :--------------------------------------------------------------: | :-----------------------------------------------------------------: | | stylegan2_config-f_ffhq_1024 | official weight | 2.8134 | 62.856/49.400 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_ffhq_1024_b4x8.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/official_weights/stylegan2-ffhq-config-f-official_20210327_171224-bce9310c.pth) | | stylegan2_config-f_lsun-car_384x512 | official weight | 5.4316 | 65.986/48.190 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_lsun-car_384x512_b4x8.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/official_weights/stylegan2-car-config-f-official_20210327_172340-8cfe053c.pth) | | stylegan2_config-f_horse_256 | official weight | - | - | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_lsun-horse_256_b4x8_800k.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/official_weights/stylegan2-horse-config-f-official_20210327_173203-ef3e69ca.pth) | | stylegan2_config-f_church_256 | official weight | - | - | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_lsun-church_256_b4x8_800k.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/official_weights/stylegan2-church-config-f-official_20210327_172657-1d42b7d1.pth) | | stylegan2_config-f_cat_256 | official weight | - | - | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_lsun-cat_256_b4x8_800k.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/official_weights/stylegan2-cat-config-f-official_20210327_172444-15bc485b.pth) | | stylegan2_config-f_ffhq_256 | our training | 3.992 | 69.012/40.417 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_ffhq_256_b4x8_800k.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/stylegan2_c2_ffhq_256_b4x8_20210407_160709-7890ae1f.pth) | | stylegan2_config-f_ffhq_1024 | our training | 2.8185 | 68.236/49.583 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_ffhq_1024_b4x8.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/stylegan2_c2_ffhq_1024_b4x8_20210407_150045-618c9024.pth) | | stylegan2_config-f_lsun-car_384x512 | our training | 2.4116 | 66.760/50.576 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_lsun-car_384x512_b4x8.py) | [model](https://download.openmmlab.com/mmgen/stylegan2/stylegan2_c2_lsun-car_384x512_b4x8_1800k_20210424_160929-fc9072ca.pth) | ## FP16 Support and Experiments Currently, we have supported FP16 training for StyleGAN2, and here are the results for the mixed-precision training. (Experiments for FFHQ1024 will come soon.) <div align="center"> <b> Evaluation FID for FP32 and FP16 training </b> <br/> <img src="https://user-images.githubusercontent.com/12726765/117523645-18e90880-afec-11eb-9327-da14362b8bfd.png" width="600"/> </div> As shown in the figure, we provide **3** ways to do mixed-precision training for `StyleGAN2`: - [stylegan2_c2_fp16_PL-no-scaler](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_fp16_partial-GD_PL-no-scaler_ffhq_256_b4x8_800k.py): In this setting, we try our best to follow the official FP16 implementation in [StyleGAN2-ADA](https://github.com/NVlabs/stylegan2-ada). Similar to the official version, we only adopt FP16 training for the higher-resolution feature maps (the last 4 stages in G and the first 4 stages). Note that we do not adopt the `clamp` way to avoid gradient overflow used in the official implementation. We use the `autocast` function from `torch.cuda.amp` package. - [stylegan2_c2_fp16-globalG-partialD_PL-R1-no-scaler](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_fp16-globalG-partialD_PL-R1-no-scaler_ffhq_256_b4x8_800k.py): In this config, we try to adopt mixed-precision training for the whole generator, but in partial discriminator (the first 4 higher-resolution stages). Note that we do not apply the loss scaler in the path length loss and gradient penalty loss. Because we always meet divergence after adopting the loss scaler to scale the gradient in these two losses. - [stylegan2_c2_apex_fp16_PL-R1-no-scaler](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_apex_fp16_PL-R1-no-scaler_ffhq_256_b4x8_800k.py): In this setting, we adopt the [APEX](https://github.com/NVIDIA/apex) toolkit to implement mixed-precision training with multiple loss/gradient scalers. In APEX, you can assign different loss scalers for the generator and the discriminator respectively. Note that we still ignore the gradient scaler in the path length loss and gradient penalty loss. | Model | Comment | Dataset | FID50k | Config | Download | | :----------------------------------------------: | :-------------------------------------: | :-----: | :----: | :-----------------------------------------------: | :-------------------------------------------------: | | stylegan2_config-f_ffhq_256 | baseline | FFHQ256 | 3.992 | [config](https://github.com/open-mmlab/mmgeneration/tree/master/configs/styleganv2/stylegan2_c2_ffhq_256_b4x8_800k.py) | [ckpt](https://download.openmmlab.com/mmgen/stylegan2/stylegan2_c2_ffhq_256_b4x8_20210407_160709-7890ae1f.pth) | | stylegan2_c2_fp16_partial-GD_PL-no-scaler_ffhq_256_b4x8_800k | partia