Add V1.3 model (#158)

* add gfpgan bilinear arch * add v1.3 * update readme * update readme * update readme * rename
TencentARC · Feb 14, 2022 · 24b1f24 · 24b1f24
1 parent c068e4d
commit 24b1f24
Show file tree

Hide file tree

Showing 9 changed files with 647 additions and 55 deletions.
diff --git a/Comparisons.md b/Comparisons.md
@@ -0,0 +1,24 @@
+# Comparisons
+
+## Comparisons among different model versions
+
+Note that V1.3 is not always better than V1.2. You may need to try different models based on your purpose and inputs.
+
+| Version | Strengths  | Weaknesses |
+| :---: | :---:        |     :---:      |
+|V1.3 |  ✓ natural outputs<br> ✓better results on very low-quality inputs <br> ✓ work on relatively high-quality inputs <br>✓ can have repeated (twice) restorations | ✗ not very sharp <br> ✗ have a slight change on identity |
+|V1.2 |  ✓ sharper output <br> ✓ with beauty makeup | ✗ some outputs are unnatural|
+
+For the following images, you may need to **zoom in** for comparing details, or **click the image** to see in the full size.
+
+| Input | V1  | V1.2 | V1.3
+| :---: | :---:        |     :---:      |  :---:      |
+|![019_Anne_Hathaway_01_00](https://user-images.githubusercontent.com/17445847/153762146-96b25999-4ddd-42a5-a3fe-bb90565f4c4f.png)|  ![](https://user-images.githubusercontent.com/17445847/153762256-ef41e749-5a27-495c-8a9c-d8403be55869.png)  | ![](https://user-images.githubusercontent.com/17445847/153762297-d41582fc-6253-4e7e-a1ce-4dc237ae3bf3.png)   | ![](https://user-images.githubusercontent.com/17445847/153762215-e0535e94-b5ba-426e-97b5-35c00873604d.png)  |
+| ![106_Harry_Styles_00_00](https://user-images.githubusercontent.com/17445847/153789040-632c0eda-c15a-43e9-a63c-9ead64f92d4a.png) | ![](https://user-images.githubusercontent.com/17445847/153789172-93cd4980-5318-4633-a07e-1c8f8064ff89.png) | ![](https://user-images.githubusercontent.com/17445847/153789185-f7b268a7-d1db-47b0-ae4a-335e5d657a18.png) | ![](https://user-images.githubusercontent.com/17445847/153789198-7c7f3bca-0ef0-4494-92f0-20aa6f7d7464.png)|
+| ![076_Paris_Hilton_00_00](https://user-images.githubusercontent.com/17445847/153789607-86387770-9db8-441f-b08a-c9679b121b85.png) | ![](https://user-images.githubusercontent.com/17445847/153789619-e56b438a-78a0-425d-8f44-ec4692a43dda.png) | ![](https://user-images.githubusercontent.com/17445847/153789633-5b28f778-3b7f-4e08-8a1d-740ca6e82d8a.png) | ![](https://user-images.githubusercontent.com/17445847/153789645-bc623f21-b32d-4fc3-bfe9-61203407a180.png)|
+| ![008_George_Clooney_00_00](https://user-images.githubusercontent.com/17445847/153790017-0c3ca94d-1c9d-4a0e-b539-ab12d4da98ff.png) | ![](https://user-images.githubusercontent.com/17445847/153790028-fb0d38ab-399d-4a30-8154-2dcd72ca90e8.png) | ![](https://user-images.githubusercontent.com/17445847/153790044-1ef68e34-6120-4439-a5d9-0b6cdbe9c3d0.png) | ![](https://user-images.githubusercontent.com/17445847/153790059-a8d3cece-8989-4e9a-9ffe-903e1690cfd6.png)|
+| ![057_Madonna_01_00](https://user-images.githubusercontent.com/17445847/153790624-2d0751d0-8fb4-4806-be9d-71b833c2c226.png) | ![](https://user-images.githubusercontent.com/17445847/153790639-7eb870e5-26b2-41dc-b139-b698bb40e6e6.png) | ![](https://user-images.githubusercontent.com/17445847/153790651-86899b7a-a1b6-4242-9e8a-77b462004998.png) | ![](https://user-images.githubusercontent.com/17445847/153790655-c8f6c25b-9b4e-4633-b16f-c43da86cff8f.png)|
+| ![044_Amy_Schumer_01_00](https://user-images.githubusercontent.com/17445847/153790811-3fb4fc46-5b4f-45fe-8fcb-a128de2bfa60.png) | ![](https://user-images.githubusercontent.com/17445847/153790817-d45aa4ff-bfc4-4163-b462-75eef9426fab.png) | ![](https://user-images.githubusercontent.com/17445847/153790824-5f93c3a0-fe5a-42f6-8b4b-5a5de8cd0ac3.png) | ![](https://user-images.githubusercontent.com/17445847/153790835-0edf9944-05c7-41c4-8581-4dc5ffc56c9d.png)|
+| ![012_Jackie_Chan_01_00](https://user-images.githubusercontent.com/17445847/153791176-737b016a-e94f-4898-8db7-43e7762141c9.png) | ![](https://user-images.githubusercontent.com/17445847/153791183-2f25a723-56bf-4cd5-aafe-a35513a6d1c5.png) | ![](https://user-images.githubusercontent.com/17445847/153791194-93416cf9-2b58-4e70-b806-27e14c58d4fd.png) | ![](https://user-images.githubusercontent.com/17445847/153791202-aa98659c-b702-4bce-9c47-a2fa5eccc5ae.png)|
+
+<!-- | ![]() | ![]() | ![]() | ![]()|  -->
diff --git a/FAQ.md b/FAQ.md
@@ -0,0 +1,7 @@
+# FAQ
+
+1. **How to finetune the GFPGANCleanv1-NoCE-C2 (v1.2) model**
+
+**A:** 1) The GFPGANCleanv1-NoCE-C2 (v1.2) model uses the *clean* architecture, which is more friendly for deploying.
+2) This model is not directly trained. Instead, it is converted from another *bilinear* model.
+3) If you want to finetune the GFPGANCleanv1-NoCE-C2 (v1.2), you need to finetune its original *bilinear* model, and then do the conversion.
diff --git a/PaperModel.md b/PaperModel.md
@@ -60,17 +60,17 @@ wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth
 - Option 1: Load extensions just-in-time(JIT)
 
     ```bash
-    BASICSR_JIT=True python inference_gfpgan.py --model_path experiments/pretrained_models/GFPGANv1.pth --test_path inputs/whole_imgs --save_root results --arch original --channel 1
+    BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1
 
     # for aligned images
-    BASICSR_JIT=True python inference_gfpgan.py --model_path experiments/pretrained_models/GFPGANv1.pth --test_path inputs/cropped_faces --save_root results --arch original --channel 1 --aligned
+    BASICSR_JIT=True python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1 --aligned
     ```
 
 - Option 2: Have successfully compiled extensions during installation
 
     ```bash
-    python inference_gfpgan.py --model_path experiments/pretrained_models/GFPGANv1.pth --test_path inputs/whole_imgs --save_root results --arch original --channel 1
+    python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1
 
     # for aligned images
-    python inference_gfpgan.py --model_path experiments/pretrained_models/GFPGANv1.pth --test_path inputs/cropped_faces --save_root results --arch original --channel 1 --aligned
+    python inference_gfpgan.py --input inputs/whole_imgs --output results --version 1 --aligned
     ```
diff --git a/README.md b/README.md
@@ -19,7 +19,11 @@
 GFPGAN aims at developing a **Practical Algorithm for Real-world Face Restoration**.<br>
 It leverages rich and diverse priors encapsulated in a pretrained face GAN (*e.g.*, StyleGAN2) for blind face restoration.
 
+:question: Frequently Asked Questions can be found in [FAQ.md](FAQ.md).
+
 :triangular_flag_on_post: **Updates**
+
+- :fire::fire::white_check_mark: Add **[V1.3 model](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)**, which produces **more natural** restoration results, and better results on *very low-quality* / *high-quality* inputs. See more in [Model zoo](#european_castle-model-zoo), [Comparisons.md](Comparisons.md)
 - :white_check_mark: Integrated to [Huggingface Spaces](https://huggingface.co/spaces) with [Gradio](https://github.com/gradio-app/gradio). See [Gradio Web Demo](https://huggingface.co/spaces/akhaliq/GFPGAN).
 - :white_check_mark: Support enhancing non-face regions (background) with [Real-ESRGAN](https://github.com/xinntao/Real-ESRGAN).
 - :white_check_mark: We provide a *clean* version of GFPGAN, which does not require CUDA extensions.
@@ -88,24 +92,54 @@ If you want to use the original model in our paper, please see [PaperModel.md](P
 
 ## :zap: Quick Inference
 
-Download pre-trained models: [GFPGANCleanv1-NoCE-C2.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth)
+We take the v1.3 version for an example. More models can be found [here](#european_castle-model-zoo).
+
+Download pre-trained models: [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth)
 
 ```bash
-wget https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth -P experiments/pretrained_models
+wget https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth -P experiments/pretrained_models
 ```
 
 **Inference!**
 
 ```bash
-python inference_gfpgan.py --upscale 2 --test_path inputs/whole_imgs --save_root results
+python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2
+```
+
+```console
+Usage: python inference_gfpgan.py -i inputs/whole_imgs -o results -v 1.3 -s 2 [options]...
+
+  -h                   show this help
+  -i input             Input image or folder. Default: inputs/whole_imgs
+  -o output            Output folder. Default: results
+  -v version           GFPGAN model version. Option: 1 | 1.2 | 1.3. Default: 1.3
+  -s upscale           The final upsampling scale of the image. Default: 2
+  -bg_upsampler        background upsampler. Default: realesrgan
+  -bg_tile             Tile size for background sampler, 0 for no tile during testing. Default: 400
+  -suffix              Suffix of the restored faces
+  -only_center_face    Only restore the center face
+  -aligned             Input are aligned faces
+  -ext                 Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto
 ```
 
 If you want to use the original model in our paper, please see [PaperModel.md](PaperModel.md) for installation and inference.
 
 ## :european_castle: Model Zoo
 
-- [GFPGANCleanv1-NoCE-C2.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth): No colorization; no CUDA extensions are required. It is still in training. Trained with more data with pre-processing.
-- [GFPGANv1.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth): The paper model, with colorization.
+| Version | Model Name  | Description |
+| :---: | :---:        |     :---:      |
+| V1.3 | [GFPGANv1.3.pth](https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.3.pth) | Based on V1.2; **more natural** restoration results; better results on very low-quality / high-quality inputs. |
+| V1.2 | [GFPGANCleanv1-NoCE-C2.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.2.0/GFPGANCleanv1-NoCE-C2.pth) | No colorization; no CUDA extensions are required. Trained with more data with pre-processing. |
+| V1 | [GFPGANv1.pth](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth) | The paper model, with colorization. |
+
+The comparisons are in [Comparisons.md](Comparisons.md).
+
+Note that V1.3 is not always better than V1.2. You may need to select different models based on your purpose and inputs.
+
+| Version | Strengths  | Weaknesses |
+| :---: | :---:        |     :---:      |
+|V1.3 |  ✓ natural outputs<br> ✓better results on very low-quality inputs <br> ✓ work on relatively high-quality inputs <br>✓ can have repeated (twice) restorations | ✗ not very sharp <br> ✗ have a slight change on identity |
+|V1.2 |  ✓ sharper output <br> ✓ with beauty makeup | ✗ some outputs are unnatural |
 
 You can find **more models (such as the discriminators)** here: [[Google Drive](https://drive.google.com/drive/folders/17rLiFzcUMoQuhLnptDsKolegHWwJOnHu?usp=sharing)], OR [[Tencent Cloud 腾讯微云](https://share.weiyun.com/ShYoCCoc)]