DEV Community

Cover image for SUPIR is Still Unchallanged Image Upscaler — Supports GPUs starting from RTX 1000 series to RTX 5000 series
Furkan Gözükara
Furkan Gözükara

Posted on

SUPIR is Still Unchallanged Image Upscaler — Supports GPUs starting from RTX 1000 series to RTX 5000 series

SUPIR is Still Unchallanged Image Upscaler — Supports GPUs starting from RTX 1000 series to RTX 5000 series including Cloud GPUs like H100, A100, B200, L40S, RTX 6000 Pro and such

by Furkan Gözükara - PhD Computer Engineer, SECourses

App Download Link

You can download SUPIR app from here : https://www.patreon.com/posts/99176057

It has 1-click installers for Windows (only Python 3.10.11 and Git should be sufficient), RunPod (official Pytorch 2.2.0 template) and Massed Compute template Creator > SECourses

App Info

  • SUPIR: Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild 1 click installer scripts.
  • SUPIR Sampler and Text CFG Comparison : https://imgsli.com/MjU2ODQz/2/1
  • Gemini 2.5 Pro prompt to get image description for free :
  • describe this image for sdxl. write single line prompt to regenerate it exactly same. make the prompt extremely detailed
  • https://aistudio.google.com/prompts/new_chat
  • Use Default preset for highest loyalty and Replicate preset for adding more details
  • Human upscale from 1024x1024 to 3072x3072 (3x upscale and total 9x resolution) with face restore comparison
  • https://imgsli.com/NDEzMDYx
  • Owl upscale from 1024x1024 to 3072x3072 (3x upscale and total 9x resolution)
  • https://imgsli.com/NDEzMDYy

Video Tutorials

Tutorials are older but hopefully a newer one will be made and they should be still useful

6 September 2025 Update V91

  • Libraries upgraded to Torch 2.8, CUDA 12.9, xFormers 0.0.33, Flash Attention 2.8.3
  • You don’t need to have CUDA or anything else installed and it should work with Python 3.10.11 and Git installed
  • When compiling libraries, I added support for all GPUs starting from RTX 1000 to 5000 series including other GPUs like A100, H100, B200, L40, etc
  • Compiled for TORCH_CUDA_ARCH_LIST=6.1;7.5;8.0;8.6;8.9;9.0;10.0;12.0
  • Downloader script is 100% improved and now it is 100x more robust even for slow connection internets
  • It has 100% continue feature + SHA256 hash check feature now
  • To upgrade libraries, extract latest zip file and overwrite files
  • Enter inside SUPIR folder and delete venv folder
  • Then run Windows_Install_or_Update.bat file
  • Installation process is also now way faster for Windows, RunPod and Massed Compute
  • To fresh install extract zip file into drive like E:\SUPIR_v91
  • Just execute Windows_Install_or_Update.bat file
  • Then once installation finished use Windows_Start_SUPIR_app.bat file
  • If you have RTX 2000 or 1000 series still should work but if you get error please report me and try If_Regular_Install_Fails_On_RTX_2000_or_Below_GPUs.bat
  • If your download of models fails for any reason, you can always run Windows_Fix_Model_Download.bat
  • It will resume and also hash check downloaded files

Windows Requirements

2 September 2025 Update V87

  • With V87 full screen image preview bug fixed
  • For update run Windows_Update_Version.bat

30 August 2025 Update V86

  • We have added the new requested features
  • Feature 1 : Skip already upscaled images in target folder option when doing batch processing
  • This is simply resume
  • Feature 2 : In Face options Only Face Restoration
  • This will apply SUPIR upscale to only face and rest of the image will be deterministic Lanczos upscale (simply exactly same when you downscale back to original resolution)
  • Feature 3 : Now when doing both batch and single processing, it will show extra screen
  • Actually this was added for batch processing but due to Gradio limitations, single processing will display too
  • This screen will be displayed at left top
  • Fix : Cancel button should work perfectly fine for Single and Batch processing now
  • With V86 from V85, messages displayed at the top panel fixed + improved
  • For update, use latest zip file, extract and overwrite, and run Windows_Update_Version.bat

27 August 2025 Update V80

  • New Feature : First Downscale to Match Target Upscale Size
  • When enabled, the image is first downscaled using high-quality Lanczos algorithm to match the target resolution (after max megapixel/resolution constraints), then upscaled by the specified factor. This can help achieve more consistent results when working with size-constrained outputs.
  • So lets say your input image is 2000x2000 = 4 mega pixels
  • You want to sharpen and increase quality with 4x upscale but that would make it 16 mega pixel which is quite not possible
  • So set Max Megapixels = 4
  • Set Upscale Size = 4
  • And enable First Downscale to Match Target Upscale Size checkbox
  • So it will first downscale image into 500x500 and then it will upscale into 2000x2000
  • To update just use latest zip file and run Windows_Update_Version.bat

How To Install

  • Download latest zip file from the attachments and extract files wherever you want to install
  • Double click windows_install .bat and let it install and download necessary models
  • Requires 12 GB VRAM without LLaVA. Tested on RTX 3060
  • Make sure that your VRAM usage is not bigger than 500 MB before starting the SUPIR
  • If you want to use LLaVA you can use our better installer : https://www.patreon.com/posts/90744385
  • LLaVA is not mandatory as shown in full tutorial video : https://youtu.be/PqREA6-bC3w LLaVA Tutorial

Massed Compute (Recommend Cloud) :

RunPod (Cloud):

How To Use

  • The biggest difference making thing is under SUPIR options Text Guidance Scale
  • This will impact how much add new details or stay loyal to the original image
  • Next big thing is Sampler
  • EDM is more loyal to the original image but DPMPP2M can add more details
  • S-Churn and S-Noise doesn’t make much valuable difference but you can test
  • In the Advanced Options
  • Linear CFG and Linear Stage2 Guidance can make difference so you can try them
  • All generated images will be automatically saved inside outputs folder
  • If you also want to use LLaVA we have better LLaVA scripts which supports 4-bit, 8-bit and 16-bit loading
  • But even with simple captions it works great you can just type
  • Image comparison sliders to Test : https://imgsli.com/ , https://web-toolbox.dev/en/tools/image-compare-slider , https://www.diffchecker.com/image-compare/
  • Here example. All default except Linear Stage2 Guidance is enabled and set 0.5
  • input image : input_1.png

Basic Caption : a trex dinosaur in jurassic park
Output : basic_caption.png

Better caption : A gigantic dinosaur with sharp teeth is standing in a lush green landscape with mountains in the background. The sky is partly cloudy, and the dinosaur appears to be in motion, possibly running or lunging forward. The setting seems to be reminiscent of prehistoric or Jurassic environments, likely intended to represent the natural habitat of such a creature.
Output : better_caption.png

Camel
Camel

A family photo
A family photo

image/png
image/png
image/png
image/jpeg
image/png
image/jpeg
image/png
image/png
image/png
image/png

Top comments (0)