Thank you :) Wanted to show you my SRFormer_light model :)
Phhofm opened this issue · 5 comments
Thank you for this network :)
I just wanted to show real quick that I have trained a (4x) SRFormer_light model for anime super-resolution with real degradation (compression, noise, blur), and created visual outputs which can be compared to a SwinIR_light (=small) model I have trained on the same dataset with the same config (so same losses, gt_size, batch size, both from scratch meaning no pretrain). Just one visual comparison (more outputs in the link below) of Input, SwinIR_light and SRFormer_light:
More visual outputs can be found here and the trained SRFormer_light model file can be found here
Greatful! The model you trained demonstrates the performance of SRFormer in anime SR. I want to express my sincere gratitude for your work.
PS I think SRFormer is impressive and its results are growing on me. I trained/finetuned another model (not lightweight but an SRFormer base model) and I think the results are very good even when compared with HAT-L (which is a bigger and slower arch/model, trained with the same config). It is trained with otf jpg compression and blur. Here are results compared with Real-ESRGAN (RRDBNet) and HAT-S and HAT-L
https://imgsli.com/MTg5MDY3/0/3
https://imgsli.com/MTg5MDYy/0/3
https://imgsli.com/MTg5MDY1/0/3
The model files (pth and onnx conversions in the onnx folder) with training info can be found here
SRFormer is accepted by ICCV 2023, We will enrich our repo recently and provide more demos, and we will add a link in our repo about the third-party model you trained so that people can learn more about SRFormer！
PS I think SRFormer is impressive and its results are growing on me. I trained/finetuned another model (not lightweight but an SRFormer base model) and I think the results are very good even when compared with HAT-L (which is a bigger and slower arch/model, trained with the same config). It is trained with otf jpg compression and blur. Here are results compared with Real-ESRGAN (RRDBNet) and HAT-S and HAT-L
https://imgsli.com/MTg5MDY3/0/3 https://imgsli.com/MTg5MDYy/0/3 https://imgsli.com/MTg5MDY1/0/3
The model files (pth and onnx conversions in the onnx folder) with training info can be found here
excelent work! can you share some info about your training dataset? that's helpful to me !
PS I think SRFormer is impressive and its results are growing on me. I trained/finetuned another model (not lightweight but an SRFormer base model) and I think the results are very good even when compared with HAT-L (which is a bigger and slower arch/model, trained with the same config). It is trained with otf jpg compression and blur. Here are results compared with Real-ESRGAN (RRDBNet) and HAT-S and HAT-L
https://imgsli.com/MTg5MDY3/0/3 https://imgsli.com/MTg5MDYy/0/3 https://imgsli.com/MTg5MDY1/0/3
The model files (pth and onnx conversions in the onnx folder) with training info can be found here
excelent work! can you share some info about your training dataset? that's helpful to me !
The dataset used for the anime pretrain is HFA2K_LUDVAE, available in the #dataset-releases channel of the Enhance Everything! Discord server.
The realistic dataset is Nomos8K, also in that Discord server.