Training on custom dataset #39

New Issue

claunia · 2026-01-29T21:38:59Z

claunia commented

2026-01-29 21:38:59 +00:00

Originally created by @3BBUAE on GitHub (Jul 28, 2021).

Hello, first of all fantastic work and thanks for sharing.

I would like to know how can I train the model on a custom dataset?
I noticed in the training explanation, there are 3 files I need to download excluding the dataset:
-Pre trained styleGAN2 model
-FFHQ component locations
-Arcface

I know that Arcface is used for face recognition. I assume that the pretrained styleGAN2 model, is to train the GFPGAN model from scratch so if i wanted to continue training I could just use the model you have provided for inference to continue training on my custom dataset. Finally for the components locations file, will it work with my custom dataset or is it specific to the FFHQ dataset, and if it won't work how will I be able to create my own so it may work with my dataset?

I hope my issue is clear, thanks.

Originally created by @3BBUAE on GitHub (Jul 28, 2021). Hello, first of all fantastic work and thanks for sharing. I would like to know how can I train the model on a custom dataset? I noticed in the training explanation, there are 3 files I need to download excluding the dataset: -Pre trained styleGAN2 model -FFHQ component locations -Arcface I know that Arcface is used for face recognition. I assume that the pretrained styleGAN2 model, is to train the GFPGAN model from scratch so if i wanted to continue training I could just use the [model ](https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth) you have provided for inference to continue training on my custom dataset. Finally for the components locations file, will it work with my custom dataset or is it specific to the FFHQ dataset, and if it won't work how will I be able to create my own so it may work with my dataset? I hope my issue is clear, thanks.

claunia closed this issue

2026-01-29 21:39:00 +00:00

claunia commented

2026-01-29 21:39:01 +00:00

@3BBUAE commented on GitHub (Jul 28, 2021):

Also does the model downsample the images while training, or will I need to provide my own version of low resolution faces?

@3BBUAE commented on GitHub (Jul 28, 2021): Also does the model downsample the images while training, or will I need to provide my own version of low resolution faces?

claunia commented

2026-01-29 21:39:03 +00:00

@MDYLL commented on GitHub (Jul 28, 2021):

I trained on my own dataset without pretrained Stylegan. Results, that I got, so-so...I didn't understand how to use my own pretrained Stylegan

@MDYLL commented on GitHub (Jul 28, 2021): I trained on my own dataset without pretrained Stylegan. Results, that I got, so-so...I didn't understand how to use my own pretrained Stylegan

claunia commented

2026-01-29 21:39:05 +00:00

@xinntao commented on GitHub (Aug 3, 2021):

@3BBUAE

You can continue your training from the pretrained models.
The FFHQ component locations are for FFHQ. For your own dataset, I think i) You can first remove the component discriminators. or ii) generate the locations with the provided scripts https://github.com/TencentARC/GFPGAN/blob/master/scripts/parse_landmark.py
The current training process only requires High-Resolution images, and the data loader will generate low-resolution ones on the fly.

@MDYLL

You can finetune from the pre-trained GFPGAN.
Or use the provided StyleGAN. (I think you may not need to train the StyleGAN on your own dataset.)

@xinntao commented on GitHub (Aug 3, 2021): @3BBUAE 1. You can continue your training from the pretrained models. 2. The FFHQ component locations are for FFHQ. For your own dataset, I think i) You can first remove the component discriminators. or ii) generate the locations with the provided scripts https://github.com/TencentARC/GFPGAN/blob/master/scripts/parse_landmark.py 3. The current training process only requires High-Resolution images, and the data loader will generate low-resolution ones on the fly. @MDYLL 1. You can finetune from the pre-trained GFPGAN. 2. Or use the provided StyleGAN. (I think you may not need to train the StyleGAN on your own dataset.)

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: TencentARC/GFPGAN#39