Releases: Phhofm/models
4xNomos8kHAT-L_bokeh_jpg
Name: 4xNomos8kHAT-L_bokeh_jpg
Author: Philip Hofmann
Release: 05.10.2023
License: CC BY 4.0
Network: HAT
Scale: 4
Purpose: 4x photo upscaler (handles bokeh effect and jpg compression)
Iterations: 145000
epoch: 66
batch_size: 4
HR_size: 128
Dataset: nomos8k
Number of train images: 8492
OTF Training: No
Pretrained_Model_G: HAT-L_SRx4_ImageNet-pretrain
Description:
4x photo upscaler, made to specifically handle bokeh effect and jpg compression. Basically a HAT-L variant of the already released 4xNomosUniDAT_bokeh_jpg model, but specifically trained for photos on the nomos8k dataset (and hopefully without the smoothing effect).
The three strengths of this model (design purpose):
Specifically for photos / photography
Handles bokeh effect
Handles jpg compression
This model will not attempt to:
Denoise
Deblur
4xNomos8kDAT
Name: 4xNomos8kDAT
Author: Philip Hofmann
Release Date: 13.08.2023
License: CC BY 4.0
Network: DAT
Architecture Option: DAT
Scale: 4
Purpose: 4x photo upscaling model for realistic sr
Iterations: 110000
epoch: 71
batch_size: 4
HR_size: 128
Dataset: Nomos8k_sfw
Number of train images: 6118
OTF Training: Yes
Pretrained_Model_G: DAT_x4.pth
Description: 4x photo upscaler with otf jpg compression, blur and resize, trained on musl's Nomos8k_sfw dataset for realisic sr, this time based on the DAT arch, as a finetune on the official 4x DAT model.
Examples:
Imgsli1 (generated with onnx file)
Imgsli2 (generated with onnx file)
Imgsli (generated with testscript of dat repo on the three test images in dataset/single with pth file)
4xLexicaDAT2_otf
Name: 4xLexicaDAT2_otf
Author: Philip Hofmann
Release: 01.11.2023
License: CC BY 4.0
Network: DAT
Scale: 4
Purpose: 4x ai generated image upscaler
Iterations: 175000
epoch: 3
batch_size: 6
HR_size: 128
Dataset: lexica
Number of train images: 43856
OTF Training: Yes
Pretrained_Model_G: DAT_2_x4.pth
Description:
4x ai generated image upscaler trained with otf
4xLSDIRDAT
Name: 4xLSDIRDAT
Author: Philip Hofmann
Release: 10.09.2023
License: CC BY 4.0
Network: DAT
Scale: 4
Purpose: 4x photo upscaler
Iterations: 95000
epoch: 2
batch_size: 4
HR_size: 128
Dataset: LSDIR
Number of train images: 81991
OTF Training: Yes
Pretrained_Model_G: DAT_x4
Description: 4x photo upscaler on the LSDIR dataset, trained with otf for jpg, resize and small blur.
4xLSDIRCompactC
Releasing two models I have been working on to extend my previous 4xLSDIRCompact model:
Name: 4xLSDIRCompactC
Author: Philip Hofmann
Release Date: 17.03.2023
License: CC BY 4.0
Network: SRVGGNetCompact
Scale: 4
Purpose: 4x photo upscaler that handler jpg compression
Iterations: 190000
batch_size: Variable(1-5)
HR_size: 256
Dataset: LSDIR
Dataset_size: 84991
OTF Training No
Pretrained_Model_G: 4xLSDIRCompact.pth
Description: Trying to extend my previous model to be able to handle compression (JPG 100-30) by manually altering the training dataset, since 4xLSDIRCompact cant handle compression. Use this instead of 4xLSDIRCompact if your photo has compression (like an image from the web).
Name: 4xLSDIRCompactR
Author: Philip Hofmann
Release Date: 17.03.2023
License: CC BY 4.0
Network: SRVGGNetCompact
Scale: 4
Purpose: 4x photo uspcaler that handles jpg compression, noise and slight
Iterations: 130000
batch_size: Variable(1-5)
HR_size: 256
Dataset: LSDIR
Dataset_size: 84991
OTF Training No
Pretrained_Model_G: 4xLSDIRCompact.pth
Description: Extending my last 4xLSDIRCompact model to Real-ESRGAN, meaning trained on synthetic data instead to handle more kinds of degradations, it should be able to handle compression, noise, and slight blur.
Here is a comparison to show that 4xLSDIRCompact cannot handle compression artifacts, and that these two models will produce better output for that specific scenario. These models are not ‘better’ than the previous one, they are just meant to handle a different use case: https://imgsli.com/MTYyODY3
4xLSDIRCompact Series 3
I am releasing the Series 3 from my 4xLSDIRCompact models. In general my suggestion is, if you have good quality input images use 4xLSDIRCompactN3, otherwise try 4xLSDIRCompactC3 which will be able to handle jpg compression and a bit of blur, or then 4xLSDIRCompactCR3, which is an interpolation between C3 and R3 to be able to handle a bit of noise additionally.
Name: 4xLSDIRCompactN3
Author: Philip Hofmann
Release Date: 11.04.2023
License: CC BY 4.0
Model Architecture: SRVGGNetCompact
Scale: 4
Purpose: Upscale good quality input photos to x4 their size
Iterations: 185'000
batch_size: Variable(1-10)
HR_size: 256
Dataset: LSDIR
Dataset_size: 84991 hr + 84991 lr
OTF Training No
Pretrained_Model_G: 4x_Compact_Pretrain.pth
Description: The original 4xLSDIRCompact a bit more trained, cannot handle degradation
Total Training Time: 32+ hours
Name: 4xLSDIRCompactC3
Author: Philip Hofmann
Release Date: 11.04.2023
License: CC BY 4.0
Model Architecture: SRVGGNetCompact
Scale: 4
Purpose: Upscale compressed photos to x4 their size
Iterations: 230’000
batch_size: Variable(1-20)
HR_size: 256
Dataset: LSDIR
Dataset_size: 8000-84991 hr& lr
OTF Training No
Pretrained_Model_G: 4xLSDIRCompact
Description: Able to handle JPG compression (30-100).
Total Training Time: 33+ hours
Name: 4xLSDIRCompactR3
Author: Philip Hofmann
Release Date: 11.04.2023
License: CC BY 4.0
Model Architecture: SRVGGNetCompact
Scale: 4
Purpose: Upscale (degraded) photos to x4 their size
Iterations: 192’500
batch_size: Variable(1-45)
HR_size: 256
Dataset: LSDIR
Dataset_size: 10000-84991 hr& lr
OTF Training No
Pretrained_Model_G: 4xLSDIRCompact.pth
Description: Trained on synthetic data, meant to handle more degradations
Total Training Time: 61+ hours
4xLSDIRCompact2
Name: 4xLSDIRCompact2
Author: Philip Hofmann
Release Date: 25.03.2023
License: CC BY 4.0
Model Architecture: SRVGGNetCompact
Scale: 4
Purpose: 4x fast photo upscaler
Iterations: CompactC 205’000 & CompactR 150’000
batch_size: Variable(1-10)
HR_size: 256
Dataset: LSDIR
Dataset_size: 84991
OTF Training No
Pretrained_Model_G: 4x_Compact_Pretrain.pth
Description: 4xLSDIRCompactv2 supersedes the previously released models, it combines all my progress on my compact model. Both CompactC and CompactR had received around 8 hours more training since release with batch size 10 (CompactR had only been up to 5 on release), and these two were then interpolated together. This allows v2 to handle some degradations, while preserving the details of the CompactC model. Examples: https://imgsli.com/MTY0Njgz/0/2
4xLSDIRCompact
Name: 4xLSDIRCompact
Author: Philip Hofmann
Release Date: 11.03.2023
License: CC BY 4.0
Model Architecture: SRVGGNetCompact
Scale: 4
Purpose: Upscale small good quality photos to 4x their size
Iterations: 160000
batch_size: Variable(1-10)
HR_size: 256
Dataset: LSDIR
Dataset_size: 84991
OTF Training No
Pretrained_Model_G: 4x_Compact_Pretrain.pth
Description: This is my first ever released self-trained sisr upscaling model 😄
15 Examples: https://imgsli.com/MTYxNDk3
4xHFA2kLUDVAESwinIR_light & 4xHFA2kLUDVAESRFormer_light
I am releasing two lightweight models for anime upscaling with realistic degradations (compression, noise, blur) trained on HFA2kLUDVAE with two different networks:
Name: 4xHFA2kLUDVAESwinIR_light
Author: Philip Hofmann
Release Date: 10.06.2023
License: CC BY 4.0
Network: SwinIR
Arch Option: SwinIR-light
Scale: 4
Purpose: An lightweight anime 4x upscaling model with realistic degradations, based on musl's HFA2k_LUDVAE dataset
Iterations: 350,000
batch_size: 3
HR_size: 256
Epoch: 99 (require iter number per epoch: 3424)
Dataset: HFA2kLUDVAE
Number of train images: 10270
OTF Training: No
Pretrained_Model_G: None
Description: 4x lightweight anime upscaler with realistic degradations (compression, noise, blur). Visual outputs can be found on https://github.com/Phhofm/models/tree/main/4xHFA2kLUDVAE_results, together with timestamps and metrics to compare inference speed on the val set with other trained models/networks on this dataset.
Name: 4xHFA2kLUDVAESRFormer_light
Author: Philip Hofmann
Release Date: 10.06.2023
License: CC BY 4.0
Network: SRFormer
Arch Option: SRFormer-light
Scale: 4
Purpose: A lightweight anime 4x upscaling model with realistic degradations, based on musl's HFA2k_LUDVAE dataset
Iterations: 350,000
batch_size: 3
HR_size: 256
Epoch: 97 (require iter number per epoch: 3424)
Dataset: HFA2kLUDVAE
Number of train images: 10270
OTF Training: No
Pretrained_Model_G: None
Description: 4x lightweight anime upscaler with realistic degradations (compression, noise, blur). Visual outputs can be found on https://github.com/Phhofm/models/tree/main/4xHFA2kLUDVAE_results, together with timestamps and metrics to compare inference speed on the val set with other trained models/networks on this dataset.
4xHFA2kLUDVAEGRL_small
Name: 4xHFA2kLUDVAEGRL_small
Author: Philip Hofmann
Release Date: 14.06.2023
License: CC BY 4.0
Network: GRL
Scale: 4
Purpose: 4x anime super-resolution with real degradation, based on musl's HFA2k_LUDVAE dataset
Iterations: 200,000
batch_size: 3
HR_size: 256
img_size: 64
Dataset: HFA2kLUDVAE
Number of train images: 10270
OTF Training: No
Pretrained_Model_G: None
Description: 4x anime super-resolution with real degradation.
Imgsli (slider comparison, same examples)
0550 : https://imgsli.com/MTg1ODky/0/3
0045 : https://imgsli.com/MTg1ODkz/0/3
