neosr is a framework for training real-world single-image super-resolution networks.

installation

Requires Python 3.11 and CUDA =>11.8

git clone https://github.com/muslll/neosr
cd neosr

Install latest Pytorch (=>2.1) and TorchVision (required).

Then install other dependencies via pip:

pip install -e .

Alternatively, use poetry (recommended on linux):

poetry install && poetry add torch@latest torchvision@latest

Note: You must use poetry shell to enter the env after installation.

(optional) If you want to convert your models (convert.py), you need to following dependencies:

pip install onnx onnxruntime-gpu onnxconverter-common onnxsim

You can also install using poetry (recommended on linux):

poetry add onnx onnxruntime-gpu onnxconverter-common onnxsim

Please read the wiki tutorial for converting your models.

quick start

Start training by running:

python train.py -opt options.yml

Where options.yml is a configuration file. Templates can be found in options. Please read the wiki Configuration Walkthrough for an explanation of each option.

features

Currently included archs:

arch	option
ESRGAN	old_esrgan
Real-ESRGAN	esrgan
SRVGGNetCompact	compact
SwinIR	swinir_small, swinir_medium
HAT	hat_s, hat_m, hat_l
OmniSR	omnisr
SRFormer	srformer_light, srformer_medium
DAT	dat_light, dat_small, dat_medium, dat_2
DITN	ditn
DCTLSA	dctlsa
SPAN	span
NLSAN	nlsan_medium, nlsan_light
DWT	dwt
EDAT	edat, edat_light
CRAFT	craft
Bicubic++	bpp, bpp_l
Real-CUGAN	cugan

Arch Inference times with provided testscript, rtx 3060

type	fps
bicubic ++	1.76
compact	1.37
span	0.92
ditn	0.76
omnisr	0.54
swinir_small	0.49
craft	0.49
srformer_light	0.43
nlsan_light	0.38
dctlsa	0.35
dat_light	0.35
esrgan	0.18
swinir_medium	0.13
dwt_light	0.12
dat_small	0.08
dat_2	0.08
dat_medium	0.07

Supported Discriminators:

net	option
U-Net SN	unet
A2-FPN	a2fpn

Supported Optimizers:

optimizer	option
Adam	Adam or adam
AdamW	AdamW or adamw
Lion	Lion or lion
LAMB	Lamb or lamb
Adan	Adan or adan

Supported models:

model	description	option
Default	Base model, supports both Generator and Discriminator	default
OTF	Builds on top of default, adding Real-ESRGAN on-the-fly degradations	otf

Supported dataset loaders:

loader	option
Paired datasets	paired
Single datasets (for inference, no GT required)	single
Real-ESRGAN on-the-fly degradation	otf

Supported losses:

loss	option
L1 Loss	L1Loss, l1
L2 Loss	MSELoss, l2
Huber Loss	HuberLoss, huber
Perceptual Loss	perceptual_opt, PerceptualLoss
GAN	gan_opt, GANLoss, MultiScaleGANLoss
YUV Color Loss	color_opt, colorloss
LDL Loss	ldl_opt
Focal Frequency	ff_opt, focalfrequencyloss

datasets

If you don't have a dataset, you can either download research datasets like DIV2K or use one of the following.

• nomos_uni (recommended): universal dataset containing real photographs and anime images
• nomos8k: dataset with real photographs only
• hfa2k: anime dataset

These datasets have been tiled and manually curated across multiple sources, including DIV8K, Adobe-MIT 5k, RAISE, FFHQ, etc.

dataset	num images	meta_info	download	sha256
nomos_uni	2989 (512x512px)	nomos_uni_metainfo.txt	GDrive (1.3GB)	6403764c3062aa8aa6b842319502004aab931fcab228f85eb94f14f3a4c224b2
nomos_uni (lmdb)	2989 (512x512px)	-	GDrive (1.3GB)	596e64ec7a4d5b5a6d44e098b12c2eaf6951c68239ade3e0a1fcb914c4412788
nomos_uni (LQ 4x)	2989 (512x512px)	nomos_uni_metainfo.txt	GDrive (92MB)	c467e078d711f818a0148cfb097b3f60763363de5981bf7ea650dad246946920
nomos_uni (LQ 4x - lmdb)	2989 (512x512px)	-	GDrive (91MB)	1d770b2c6721c97bd2679db68f43a9f12d59a580e9cfeefd368db5a4fab0f0bb
nomos8k	8492 (512x512px)	nomos8k_metainfo.txt	GDrive (3.4GB)	89724f4adb651e1c17ebee9e4b2526f2513c9b060bc3fe16b317bbe9cd8dd138
hfa2k	2568 (512x512px)	hfa2k_metainfo.txt	GDrive (3.2GB)	3a3d2293a92fb60507ecd6dfacd636a21fd84b96f8f19f8c8a55ad63ca69037a

Note: these are not intended for use in academic research.

community datasets

These are datasets made by the upscaling community. More info can be found in the Enhance Everything discord

• kim's 8k Dataset V2: Video Game Dataset

• FaceUp: Curated version of FFHQ

• SSDIR: Curated version of LSDIR.

dataset	num images	meta_info	download	sha256
@Kim2091's 8k Dataset V2	672 (7680x4320px)	-	GDrive (33.5GB)	-
@Phhofm FaceUp	10000 (512x512)	-	GDrive (4GB)	-
@Phhofm SSDIR	10000 (512x512)	-	Gdrive (4.5GB)	-

resources

• OpenModelDB
• chaiNNer
• Training Guide from @Sirosky
• Training Info from @Kim

support me

☕ Consider supporting me on KoFi. ☕

license and acknowledgements

Released under the Apache license. This code was originally based on BasicSR. See other licenses in license/readme.

Thanks to victorca25/traiNNer, styler00dollar/Colab-traiNNer and timm for providing helpful insights into some problems.

Thanks to contributors @Phhofm, @Sirosky, @Kim2091 and @terrainer for helping with tests and bug reporting.