Models

Try out most of my models online here

This is a repo for me to publish my self trained single image super resolution (sisr) models.

Update 08.04.2024:
I am currently in the process of updating this repo by creating a github release on this repo for each released model, check out the Released Section in this partly updated README. Then I can also remove most file folders in this repo.

As an example, you can have a look at my 4xRealWebPhoto_v3_atd release which you can try out currently with chaiNNer nightly, or then the dat2 model version of this with normal chaiNNer.

For convenience, you can download a zip file of all my 111 released models (mostly as safetensors files) here.

After releasing all models here as github releases, I will also release them on Hugging Face so they are automatically downloadable if used in an application, or used in a huggingface space for example, which i had made two just to showcase, youll find them in the link.

I recommend running these models locally with chaiNNer. I made a youtube video on how to set up and use chaiNNer. Btw they externalized their upscaling code into the Spandrel Library

Results for some of these can be compared on my interactive visual comparison website, but this site is currently not up-to-date since it became a 70GB repo so automatic deployment became unfeasible and updating a bit more cumbersome, maybe I will be able in the future to reduce it a bit with BFG Repo-Cleaner.

I also made some youtube videos you might find interesting, like this one.

Find more sisr models trained by the community in openmodeldb.

Here also a link to the Enhance Everything! Discord Server where I had been active.

Also this weekend I played around with a comfyui workflow using SUPIR, I just uploaded the result in the SUPIR folder. Basically diffusion based upscalers in general have the tendency to produce output that is pretty different to an input image as to resemble more an 'img2img enlarger' process than super resolving. My play around was to try to use the consistency of transformers as the upscale, and then SUPIR in the second step but with settings that enforce consistency. Use case would be for very degraded input image where my transformer model hits a limit. Examples and readme and workflow in the folder.

Released (sorted by new)

-- Newly Updated Section --

All my released models simply downloadable through a python script

Model releases sorted by date, linked to their github release:
14.10.2025 - PSISRD_val125 (a cc0 validation set, more info in the huggingface dataset card)
14.10.2025 - 1xgaterv3_r_sharpen
14.10.2025 - 1xgaterv3_r_restore
14.10.2025 - 2xBHI_small_esrgan_pretrain
14.10.2025 - 2xBHI_small_drct-xl_pretrain
23.06.2025 - 2xPublic_realplksr_dysample_layernorm_real
23.06.2025 - 2xPublic_realplksr_dysample_layernorm_real_nn
23.06.2025 - 2xPublic_realplksr_dysample_layernorm_gan
23.06.2025 - 2xPublic_realplksr_dysample_layernorm_pretrain
21.05.2025 - 2xBHI_small_realplksr_large_dysample_pretrain
21.05.2025 - 2xBHI_small_realplksr_large_pretrain
21.05.2025 - 2xBHI_small_realplksr_dysample_pretrain
21.05.2025 - 2xBHI_small_realplksr_pretrain
21.05.2025 - 2xBHI_small_realplksr_small_dysample_pretrain
21.05.2025 - 2xBHI_small_realplksr_small_pretrain
19.05.2025 - 2xBHI_small_esc_large_pretrain
19.05.2025 - 2xBHI_small_esc_pretrain
19.05.2025 - 2xBHI_small_esc_light_pretrain
19.05.2025 - 2xBHI_small_span_pretrain
19.05.2025 - 2xBHI_small_span_fast_pretrain
17.01.2025 - 2xBHI_small_realplksr_dysample_pretrain (2x realplksr bicubic only pretrain)
12.01.2025 - 2xBHI_small_compact_pretrain (2x compact bicubic only pretrain, goal was to reach high metrics on psnry)
27.12.2024 - 4xBHI_dat2_otf_nn (4x dat2 model, similiar to 4xBHI_dat2_otf but trained without noise. Handles resizes, and jpg compression)
14.12.2024 - 4xBHI_dat2_real (4x dat2 model for web images/real degraded photos with (but not only) lens blur, realistic noise and jpeg&webp (re)compression handling)
14.12.2024 - 4xBHI_dat2_otf (4x dat2 model trained with real-esrgan otf pipeline on my bhi dataset)
14.12.2024 - 4xBHI_dat2_multiblurjpg & 4xBHI_dat2_multiblur (4x dat2 model on my bhi dataset, trained on multiscaled LR and jpg compression)
14.12.2024 - 4x BHI realplksr models (4x models with realplksr dysample on my bhi dataset: multi, multiblur, otf_nn, otf and real)
14.12.2024 - My BHI SISR dataset, consisting of around 390k training tiles, 512x512, webp format - external huggingface link.
11.11.2024 - 4xpix_sebica_pretrains (Two simple sebica pretrains of 100k iters for both sebica and sebica_mini options)
07.11.2024 - DF2K_BHI Dataset (A BHI filtered version of DF2K, which I released together with my huggingface community blog post on my BHI filtering method)
21.09.2024 - 2xAoMR_mosr (A 2x game texture mosr upscaling model, trained on and for (but not limited to) Age of Mythology: Retold textures)
05.09.2024 - 4xNomos2_hq_drct-l (A 4x drct-l upscale model for good quality/non-degraded input)
05.09.2024 - 4xNomos2_hq_atd (A 4x atd upscale model for good quality/non-degraded input)
29.08.2024 - 4xNomos2_hq_dat2 (A 4x dat2 upscale model for good quality/non-degraded input)
25.08.2024 - 4xNomos2_hq_mosr (A 4x mosr upscale model for good quality/non-degraded input)
25.08.2024 - 4xmssim_mosr_pretrain (A 4x mssim only mosr pretrain model at 420k iters for training mosr models)
16.08.2024 - 4xNomosRealWeb Dataset (Degraded x4 LR version with realistic degradations of the Nomosv2 dataset for training 4x upscaling models for upscaling images downloaded from the web).
16.08.2024 - SSDIR Dataset (Curated version of the LSDIR dataset, for the purpose of training single image super resolution models).
16.08.2024 - Lexica Dataset (Dataset for the purpose of training single image super resolution models for AI generated images).
16.08.2024 - FaceUp Dataset (Curated version of the [FFHQ dataset](https://github.com/NV