Nice to release a model again! This one is intended for realistic imagery, and works especially well on faces, hair, and nature shots. It should only be used on somewhat clear shots, without a lot of grain. I trained this model on SPAN, which as of the time of release, you'll need chaiNNer-nightly for. I aimed for a softer, more natural look for this model with as few artifacts as possible.

In addition to the Normal model, I've included a "soft" model. The Soft model is... softer. Basically it was an earlier version of the model with a more limited dataset. It produces more natural output on games or rendered content, but suffers a bit more with realistic stuff.

Note: In shots with DOF (depth of field) or bokeh, unfortunately there will be artifacts. 

Compatibility: You'll have to use the latest chaiNNer-nightly to use this model

ClearRealityV1

A 4x photo upscale DAT model trained with otf (resize, jpg, small blur) on the LSDIR dataset.

4xLSDIRDAT

Pretrained: Manga109Attempt

The Misc model is trained on various pictures shot by myself (Alsa), including bricks, stone, dirt, grass, plants, wood, bark, metal and a few others.

Realistic Misc

[Link to Github Release](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_atd)

4xNomos2_hq_atd  
Scale: 4  
Architecture: [ATD](https://github.com/LabShuHangGU/Adaptive-Token-Dictionary)  
Architecture Option: [atd](https://github.com/muslll/neosr/blob/dc4e3742132bae2c2aa8e8d16de3a9fcec6b1a74/neosr/archs/atd_arch.py#L891)  

Author: Philip Hofmann  
License: CC-BY-0.4  
Purpose: Upscaler  
Subject: Photography  
Input Type: Images  
Release Date: 05.09.2024  

Dataset: [nomosv2](https://github.com/muslll/neosr/?tab=readme-ov-file#-datasets)  
Dataset Size: 6000  
OTF (on the fly augmentations): No  
Pretrained Model: 003_ATD_SRx4_finetune  
Iterations: 180'000  
Batch Size: 2  
Patch Size: 48  
Norm: true  

Description:   
An atd 4x upscaling model, similiar to the [4xNomos2_hq_dat2](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_dat2) or [4xNomos2_hq_mosr](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr) models, trained and for usage on non-degraded input to give good quality output.

4xNomos2_hq_atd

[Link to Github Release](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_dat2)

# 4xNomos2_hq_dat2   
Scale: 4   
Architecture: [DAT](https://github.com/zhengchen1999/dat)   
Architecture Option: [dat2](https://github.com/muslll/neosr/blob/5fba7f162d36052010169e6517dec3b406c569ab/neosr/archs/dat_arch.py#L1111)   

Author: Philip Hofmann   
License: CC-BY-0.4   
Purpose: Upscaler   
Subject: Photography   
Input Type: Images   
Release Date: 29.08.2024   

Dataset: [nomosv2](https://github.com/muslll/neosr/?tab=readme-ov-file#-datasets)   
Dataset Size: 6000   
OTF (on the fly augmentations): No   
Pretrained Model: DAT_2_x4   
Iterations: 140'000   
Batch Size: 4   
Patch Size: 48   

Description:    
A dat2 4x upscaling model, similiar to the [4xNomos2_hq_mosr](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr) model, trained and for usage on non-degraded input to give good quality output.    

I scored 7 validation outputs of each of the 21 checkpoints (10k-210k) of this model training with 68 metrics.    
[The metric scores can be found in this google sheet](https://docs.google.com/spreadsheets/d/1NL-by7WvZyDMHj5XN8UeDALVSSwH70IKvwV65ATWqrA/edit?usp=sharing).   
The corresponding image files for this scoring can be [found here](https://drive.google.com/file/d/1ZTp9fBMeawftNqzg4RN9_zIvHtul5jVc/view?usp=sharing)     
Screenshot of the google sheet:     
![|100](https://i.slow.pics/VZJTsrUv.webp)

Release checkpoint has been selected by looking at the scores, manually inspecting, and then getting responses on discord to this quick visual test, A B or C, which denote different checkpoints: https://slow.pics/c/8Akzj6rR   

Checkpoint B is the one released here, but you can also try out [Checkpoint A](https://github.com/Phhofm/models/releases/download/4xNomos2_hq_dat2/4xNomos2_hq_dat2_150000.pth) or [Checkpoint C](https://github.com/Phhofm/models/releases/download/4xNomos2_hq_dat2/4xNomos2_hq_dat2_10000.pth) if you like them better.

## Model Showcase:
[Slowpics](https://slow.pics/c/yuue9WpF)

4xNomos2_hq_dat2

[Link to Github Release](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_drct-l)

4xNomos2_hq_drct-l  
Scale: 4  
Architecture: [DRCT](https://github.com/ming053l/DRCT)  
Architecture Option: [drct_l](https://github.com/muslll/neosr/blob/dc4e3742132bae2c2aa8e8d16de3a9fcec6b1a74/neosr/archs/drct_arch.py#L937)  

Author: Philip Hofmann  
License: CC-BY-0.4  
Purpose: Upscaler  
Subject: Photography  
Input Type: Images  
Release Date: 08.09.2024  

Dataset: [nomosv2](https://github.com/muslll/neosr/?tab=readme-ov-file#-datasets)  
Dataset Size: 6000  
OTF (on the fly augmentations): No  
Pretrained Model: DRCT-L_X4  
Iterations: 200'000  
Batch Size: 2  
Patch Size: 64  

Description: 
An drct-l 4x upscaling model, similiar to the [4xNomos2_hq_atd](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_atd), [4xNomos2_hq_dat2](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_dat2) and [4xNomos2_hq_mosr](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr) models, trained and for usage on non-degraded input to give good quality output.   

Model Showcase:
[Slowpics](https://slow.pics/c/1mODuuUS)

4xNomos2_hq_drct-l

[Link to Github Release](https://github.com/Phhofm/models/releases/tag/4xNomos2_hq_mosr)

# 4xNomos2_hq_mosr  
Scale: 4  
Architecture: [MoSR](https://github.com/umzi2/MoSR)  
Architecture Option: [mosr](https://github.com/umzi2/MoSR/blob/95c5bf73cca014493fe952c2fbc0bdbe593da08f/neosr/archs/mosr_arch.py#L117)  

Author: Philip Hofmann  
License: CC-BY-0.4  
Purpose: Upscaler  
Subject: Photography  
Input Type: Images  
Release Date: 25.08.2024  

Dataset: [nomosv2](https://github.com/muslll/neosr/?tab=readme-ov-file#-datasets)  
Dataset Size: 6000  
OTF (on the fly augmentations): No  
Pretrained Model: [4xmssim_mosr_pretrain](https://github.com/Phhofm/models/releases/tag/4xmssim_mosr_pretrain)  
Iterations: 190'000  
Batch Size: 6  
Patch Size: 64  

Description:   
A 4x [MoSR](https://github.com/umzi2/MoSR) upscaling model, meant for non-degraded input, since this model was trained on non-degraded input to give good quality output.   

If your input is degraded, use a 1x degrade model first. So for example if your input is a .jpg file, you could use a 1x dejpg model first.  

Model Showcase:  [Slowpics](https://slow.pics/c/cqGJb0gT)

4xNomos2_hq_mosr

[Link to Github Release](https://github.com/Phhofm/models/releases/4xNomos8k_atd_jpg)

Name: 4xNomos8k_atd_jpg   
License: CC BY 4.0   
Author: Philip Hofmann   
Network: [ATD](https://github.com/LabShuHangGU/Adaptive-Token-Dictionary)   
Scale: 4   
Release Date: 22.03.2024   
Purpose: 4x photo upscaler, handles jpg compression   
Iterations: 240'000   
epoch: 152   
batch_size: 6, 3   
HR_size: 128, 192   
Dataset: nomos8k   
Number of train images: 8492   
OTF Training: Yes   
Pretrained_Model_G: 003_ATD_SRx4_finetune   

Description:
4x photo upscaler which handles jpg compression. This model will preserve noise. Trained on the very recently released (~2 weeks ago) Adaptive-Token-Dictionary network.   

Training details: 
AdamW optimizer with U-Net SN discriminator and BFloat16.
Degraded with otf jpg compression down to 40, re-compression down to 40, together with resizes and the blur kernels.  
Losses: PixelLoss using CHC (Clipped Huber with Cosine Similarity Loss), PerceptualLoss using Huber, GANLoss, [LDL](https://github.com/csjliang/LDL) using Huber, YCbCr Color Loss (bt601) and Luma Loss (CIE XYZ) on [neosr](https://github.com/muslll/neosr).

7 Examples:
[Slowpics](https://slow.pics/s/uwnoI435)

 4xNomos8k_atd_jpg

I release my span otf series:
They come in three variations: weak, middle, and strong.
Mainly meant for photos (can be tried on other things of course).

(Also there is a non-otf span model I had been working on simultaneously that I will release shortly, should give better results on less degraded input in comparison to this span otf series)

Basically I trained the otf_strong for 90k iter and then medium and weak based off that, with some more training to de-learn (tone down) the (too?) strong degradations. Used discrim resets to correct occuring colorloss in all of them. gt size was for the most part 512 with batch 9 (since i hoped it would give better results) with 0.55 it/s training speed (first 40k at the beginning were gt size 256 with batch 20 with 0.58 it/s).

Nomos8k_span_otf_medium

Nomos8k_span_otf_strong

Nomos8k_span_otf_weak

Nomos8kHAT-L_otf

NomosUniDAT_otf

4x general purpose upscaler for non-degraded content.

Has been trained as paired dataset on box downsamples as a DAT2 model, AdamW with pixel, perceptual, gan, color, ldl and ff loss.

4xNomosUniDAT2_box

[Omni Aggregation Networks for Lightweight Image Super-Resolution (OmniSR)](https://github.com/Francis0625/Omni-SR)

While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional
self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to
include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks,
this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves recordhigh performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR

OmniSR 4x DF2K

OmniSR 4x DIV2K

Purpose: upscaling model for amateur to professional photos (regular)

Skilled in working with cats, hair, parties, and creating clear images. Also proficient in resizing photos and enlarging large, sharp images. Can effectively improve images from small sizes as well (300px at smallest on one side, depending on the subject). Experienced in experimenting with techniques like upscaling with this model twice and then reducing it by 50% to enhance details, especially in features like hair or animals.

Original image, second is upscale with the model once, and the last is upscaled with model twice.

PurePhoto span

[Link to Github Release](https://github.com/Phhofm/models/releases/4xRealWebPhoto_v2_rgt_s)

Name: 4xRealWebPhoto_v2_rgt_s   
License: CC BY 4.0   
Author: Philip Hofmann   
Network: RGT   
Network Option: RGT-S   
Scale: 4   
Release Date: 10.03.2024   
Purpose: 4x real web photo upscaler, meant for upscaling photos downloaded from the web   
Iterations: 220'000   
epoch: 5   
batch_size: 16   
HR_size: 128   
Dataset: 4xRealWebPhoto_v2 (see details in attached pdf file in github release)   
Number of train images: 1'086'976 (or 543'488 pairs)   
OTF Training: No   
Pretrained_Model_G: RGT_S_x4   

Description:
4x real web photo upscaler, meant for upscaling photos downloaded from the web. Trained on my v2 of my 4xRealWebPhoto dataset, it should be able to handle realistic noise, jpg and webp compression and re-compression, scaling and rescaling with multiple downscampling algos, and handle a little bit of lens blur.

Thought featuring degraded images in the examples, this model should also be able to handle good quality input.

Details about the approach/dataset I made to train this model (and therefore also what this model would be capable of handling) is in the attached pdf in the github release.

My previous tries of this dataset, meaning v0 and v1, will get a separate entry, though this version would be recommended over them.

12 Examples on [Slowpics](https://slow.pics/s/HxXtE1eZ)

 4xRealWebPhoto_v2_rgt_s

[Link to Github Release](https://github.com/Phhofm/models/releases/4xRealWebPhoto_v3_atd)

Name: 4xRealWebPhoto_v3_atd   
License: CC BY 4.0   
Author: Philip Hofmann   
Network: [ATD](https://github.com/LabShuHangGU/Adaptive-Token-Dictionary)   
Scale: 4   
Release Date: 22.03.2024   
Purpose: 4x upscaler for photos downloaded from the web   
Iterations: 250'000   
epoch: 10   
batch_size: 6, 3   
HR_size: 128, 192   
Dataset: 4xRealWebPhoto_v3   
Number of train images: 101'904   
OTF Training: No   
Pretrained_Model_G: 003_ATD_SRx4_finetune   

Description:   
4x real web photo upscaler, meant for upscaling photos downloaded from the web. Trained on my v3 of my 4xRealWebPhoto dataset, it should be able to handle noise, jpg and webp (re)compression, (re)scaling, and just a little bit of lens blur, while also be able to handle good quality input. Trained on the very recently released (~2 weeks ago) Adaptive-Token-Dictionary network. 

My 4xRealWebPhoto dataset tried to simulate the use-case of a photo being uploaded to the web and being processed by the service provides (like on a social media platform) so compression/downscaling, then maybe being downloaded and re-uploaded by another used where it, again, were processed by the service provider. I included different variants in the dataset. The pdf with info to the v2 dataset can be found [here](https://github.com/Phhofm/models/releases/download/4xRealWebPhoto_v2_rgt_s/4xRealWebPhoto_v2.pdf), while i simply included whats different in the v3 png:

![4xRealWebPhoto_v3](https://github.com/Phhofm/models/assets/14755670/2ec67e48-bf21-4b57-9f27-69bc49b84315)


Training details: 
AdamW optimizer with U-Net SN discriminator and BFloat16.
Degraded with otf jpg compression down to 40, re-compression down to 40, together with resizes and the blur kernels.  
Losses: PixelLoss using CHC (Clipped Huber with Cosine Similarity Loss), PerceptualLoss using Huber, GANLoss, [LDL](https://github.com/csjliang/LDL) using Huber, [Focal Frequency](https://github.com/EndlessSora/focal-frequency-loss), [Gradient Variance](https://github.com/lusinlu/gradient-variance-loss) with Huber, YCbCr Color Loss (bt601) and Luma Loss (CIE XYZ) on [neosr](https://github.com/muslll/neosr) with norm: true.

11 Examples:
[Slowpics](https://slow.pics/s/plgWVh4j)

 4xRealWebPhoto_v3_atd

[Link to Github Release](https://github.com/Phhofm/models/releases/4xRealWebPhoto_v4_dat2)

## 4xRealWebPhoto_v4_dat2

**Scale:** 4   
**Architecture:** DAT   

**Author:** Philip Hofmann   
**License:** CC-BY-4.0   
**Purpose:** Compression Removal, Deblur, Denoise, JPEG, WEBP, Restoration   
**Subject:** Photography   
**Input Type:** Images   
**Date:** 04.04.2024   

**Architecture Option:** DAT-2   
**I/O Channels:** 3(RGB)->3(RGB)   

**Dataset:** Nomos8k   
**Dataset Size:** 8492   
**OTF (on the fly augmentations):** No   
**Pretrained Model:** DAT_2_x4   
**Iterations:** 243'000   
**Batch Size:** 4-6   
**GT Size:** 128-256   

**Description:** 4x Upscaling Model for Photos from the Web. The dataset consists of only downscaled photos (to handle good quality), downscaled and compressed photos (uploaded to the web and compressed by service provider), and downscale, compressed, rescaled, recompressed photos (downloaded from the web and re-uploaded to the web).

Applied lens blur, realistic noise with my ludvae200 model, JPG and WEBP compression (40-95), and down_up, linear, cubic_mitchell, lanczos, gaussian and box downsampling algorithms. For details on the degradation process, check out the pdf with its explanations and visualizations.

This is basically a dat2 version of my previous 4xRealWebPhoto_v3_atd model, but trained with a bit stronger noise values, and also a single image per variant so drastically reduced training dataset size.
 

**Showcase:** 
[12 Slowpics Examples](https://slow.pics/s/TvJ21pJG)

4xRealWebPhoto_v4_dat2

Rybu

Purpose: Intended as a pretrain model, can be used to upscale undegraded photos.

Original Link: https://github.com/terrainer/AI-Upscaling-Models/tree/main/4xSPANkendata

4x realistic upscaler that may also work for general purpose usage. Trained on Bicubic downscaled tiles from very high quality images.

Iterations: 500k, fine-tuned with 120k, then 110k, then 75k

batch_size: 32, 16, 12, 10

HR_size: 192, 256, 320, 192

Epoch: Unknown, 32, 20, 12

Dataset_size: 58,420 512x512 tiles, downscaled with a variety of traditional algorithms and CAR

SPANkendata

[Structure-Preserving Super Resolution with Gradient Guidance](https://github.com/Maclory/SPSR)

Structures matter in single image super resolution (SISR). Recent studies benefiting from generative adversarial network (GAN) have promoted the development of SISR by recovering photo-realistic images. However, there are always undesired structural distortions in the recovered images. In this paper, we propose a structure-preserving super resolution method to alleviate the above issue while maintaining the merits of GAN-based methods to generate perceptual-pleasant details. Specifically, we exploit gradient maps of images to guide the recovery in two aspects. On the one hand, we restore high-resolution gradient maps by a gradient branch to provide additional structure priors for the SR process. On the other hand, we propose a gradient loss which imposes a second-order restriction on the super-resolved images. Along with the previous image-space loss functions, the gradient-space objectives help generative networks concentrate more on geometric structures. Moreover, our method is model-agnostic, which can be potentially used for off-the-shelf SR networks. Experimental results show that we achieve the best PI and LPIPS performance and meanwhile comparable PSNR and SSIM compared with state-of-the-art perceptual-driven SR methods. Visual results demonstrate our superiority in restoring structures while generating natural SR images.

4x SPSR

Real-World Super-Resolution via Kernel Estimation and Noise Injection

Our solution is the winner of CVPR NTIRE 2020 Challenge on Real-World Super-Resolution in both tracks.

Recent state-of-the-art super-resolution methods have achieved impressive performance on ideal datasets regardless of blur and noise. However, these methods always fail in real-world image super-resolution, since most of them adopt simple bicubic downsampling from high-quality images to construct Low-Resolution (LR) and High-Resolution (HR) pairs for training which may lose track of frequency-related details. To address this issue, we focus on designing a novel degradation framework for real-world images by estimating various blur kernels as well as real noise distributions. Based on our novel degradation framework, we can acquire LR images sharing a common domain with real-world images. Then, we propose a real-world super-resolution model aiming at better perception. Extensive experiments on synthetic noise data and real-world images demonstrate that our method outperforms the state-of-the-art methods, resulting in lower noise and better visual quality. In addition, our method is the winner of NTIRE 2020 Challenge on both tracks of Real-World Super-Resolution, which significantly outperforms other competitors by large margins.

for corrupted images with processing noise.

RealSR DF2K

Real-World Super-Resolution via Kernel Estimation and Noise Injection

Our solution is the winner of CVPR NTIRE 2020 Challenge on Real-World Super-Resolution in both tracks.

Recent state-of-the-art super-resolution methods have achieved impressive performance on ideal datasets regardless of blur and noise. However, these methods always fail in real-world image super-resolution, since most of them adopt simple bicubic downsampling from high-quality images to construct Low-Resolution (LR) and High-Resolution (HR) pairs for training which may lose track of frequency-related details. To address this issue, we focus on designing a novel degradation framework for real-world images by estimating various blur kernels as well as real noise distributions. Based on our novel degradation framework, we can acquire LR images sharing a common domain with real-world images. Then, we propose a real-world super-resolution model aiming at better perception. Extensive experiments on synthetic noise data and real-world images demonstrate that our method outperforms the state-of-the-art methods, resulting in lower noise and better visual quality. In addition, our method is the winner of NTIRE 2020 Challenge on both tracks of Real-World Super-Resolution, which significantly outperforms other competitors by large margins.

for real images taken by cell phone camera.

Architecture	SPAN
Scale	4x
Size	48nf
Color Mode	RGB
License	CC-BY-SA-4.0 Private use Commercial use Distribution Modifications Credit required Same License State Changes No Liability & Warranty Disclaimer
Date	2024-02-08
Dataset	6500 handpicked photos
Training iterations	488000
Training batch size	4
Training HR size	256
Training OTF	No

PurePhoto span

Similar Models