I'm looking at the AI scene with interest and there are some intriguing sites out there. Does anyone use image upscaling to any great effect? Dabbling, I can see that it can be of use when upscaling web-spec images reasonably well at twice the original size and it gets a little ropier as you go up the scale. However, I've looked at upscaling some already good quality print images and pretty much without exception across the sites I've tried, there has been notable colour shifting. Is it because they are web-based, colour profiles are being ignored? Admittedly I'm just looking at the free versions of each, but how viable could these be as a commercial tool?
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AI image upscaling
- Thread starter dazzer21-2
- Start date
- Sort by reaction score
Yes, I'm using image upscaling occasionally. Number one reason is Large Format Printing (LPF), if good quality images need to get upscaled. Then there are images with poor resolutions that need to get printed in decent sizes. Third, there are sometimes situations when an image or a part of an image needs more pixel for web or video.
In many cases, especially if scaling ≤150%, the standard image editing tools are good enough to get this done. Especially if scaling ≥150% it's worth to use a dedicated tool.
Must it be AI? It's hard to answer as AI is evolving fast. By now my favorite image enlarger is still non AI. If you ask for free tools, your best non AI solution should be SmillaEnlarger. What I still often use, is the similar commercial tool BenVista PhotoZoom Pro.
I'm not aware of good quality free AI image resize tools, but I'm experimenting with ON1 Resize AI, Skylum's Luminar Plugin Upscale AI and in the future with Topaz Photo AI upscaling. Results are varying and dependent on the very individual image. You're right that colors are shifting in most cases. As color correction is one of my last editing step, usually I don't care about shifting colors so much. What free AI resize tools did you try so far? How do they compare to each other and to traditional image resizing?
Keep in mind that image quality is not just resulting from upscaling, but also from processes like RAW developing, denoising, sharpening, debluring, demosaicing or dehazeing and most important a good quality photograph, rendering or scan as a basis for further editing.
In many cases, especially if scaling ≤150%, the standard image editing tools are good enough to get this done. Especially if scaling ≥150% it's worth to use a dedicated tool.
Must it be AI? It's hard to answer as AI is evolving fast. By now my favorite image enlarger is still non AI. If you ask for free tools, your best non AI solution should be SmillaEnlarger. What I still often use, is the similar commercial tool BenVista PhotoZoom Pro.
I'm not aware of good quality free AI image resize tools, but I'm experimenting with ON1 Resize AI, Skylum's Luminar Plugin Upscale AI and in the future with Topaz Photo AI upscaling. Results are varying and dependent on the very individual image. You're right that colors are shifting in most cases. As color correction is one of my last editing step, usually I don't care about shifting colors so much. What free AI resize tools did you try so far? How do they compare to each other and to traditional image resizing?
Keep in mind that image quality is not just resulting from upscaling, but also from processes like RAW developing, denoising, sharpening, debluring, demosaicing or dehazeing and most important a good quality photograph, rendering or scan as a basis for further editing.
upscayl
https://github.com/upscayl/upscayl
FreeScaler
https://github.com/TheMurusTeam/FreeScaler
Final2x
https://github.com/Tohrusky/Final2x
Drovp
https://github.com/drovp/drovp/
https://github.com/upscayl/upscayl
FreeScaler
https://github.com/TheMurusTeam/FreeScaler
Final2x
https://github.com/Tohrusky/Final2x
Drovp
https://github.com/drovp/drovp/
We bought Topaz AI Photo for a large book job that was dealing with a lot of old and badly scanned images. It worked OK, not as well as I would have liked across the board. I would certainly shop around more, we jumped on it because it did a great job in the demo on one image, but made the mistake of not testing it out on a few different pictures.
For photo restoration: GFPGAN and "Other recommended projects"
https://github.com/TencentARC/GFPGAN
https://github.com/xinntao/Real-ESRGAN
https://github.com/XPixelGroup/BasicSR
https://github.com/TencentARC/GFPGAN
https://github.com/xinntao/Real-ESRGAN
https://github.com/XPixelGroup/BasicSR
I ordered the entire Topaz AI suite during a bundle sale a year (or two?) ago. Works great for my needs and FAR superior to products I've tried previously starting with Noise Ninja way back in the PowerPC era.
The only Topaz component I've found kind of useless is the video processor.
The only Topaz component I've found kind of useless is the video processor.
I bought the Gigapixel version 5.4.5 few years back and it still works great. Never needed to upgrade to 7 or even 6.
I tried the new version 7. Results were no better than version 5.4.5. So i did not upgrade.
Also for some reason CPU processing gives better results than any GPU processing on Gigapixel.
I tried the new version 7. Results were no better than version 5.4.5. So i did not upgrade.
Also for some reason CPU processing gives better results than any GPU processing on Gigapixel.
I've found Pixelmator's built-in "Super Resolution" works pretty darn well for print quality images.
It's been my experience that most of these image upscalers work fairly well with anime type images, or photos with "patches" of color and/or line-art. But none of them work great on photos with fine details.
It's been my experience that most of these image upscalers work fairly well with anime type images, or photos with "patches" of color and/or line-art. But none of them work great on photos with fine details.
Another alternative is Krita with its AI plugin. It includes upscaling with various AI models and refining the output with additional Stable Diffusion models to add details where there were none before.
Also handy for photo restoration purposes. Generative AI in Krita works (much) better than Photoshop, which is rather limited compared.
Completely free too. A good NVIDIA CUDA GPU works best, but it also works on the Studio M2 Macs at work for me. Not the fastest, but completely doable. Also, did I mention free and (far) better results than Photoshop or Adobe's Firefly? And no limits in prompting? And extending it with any AI model and LoRA you want through the Civitai website? Full freedom?
Get it now at krita.org and download the plugin at https://github.com/Acly/krita-ai-diffusion
Then follow a few tutorials. Look for those on Youtube.
Also handy for photo restoration purposes. Generative AI in Krita works (much) better than Photoshop, which is rather limited compared.
Completely free too. A good NVIDIA CUDA GPU works best, but it also works on the Studio M2 Macs at work for me. Not the fastest, but completely doable. Also, did I mention free and (far) better results than Photoshop or Adobe's Firefly? And no limits in prompting? And extending it with any AI model and LoRA you want through the Civitai website? Full freedom?
Get it now at krita.org and download the plugin at https://github.com/Acly/krita-ai-diffusion
Then follow a few tutorials. Look for those on Youtube.
I don't understand where should I unzip the plugin.
The ReadMe file says "MacOS: Usually ~/Library/Application Support/krita/pykrita"
The latest release says "MacOS: ~/Library/Application Support/krita/ai_diffusion"
Neither work.
Edit: I got it to work
- Unzip in ~/Library/Application Support/krita/pykrita
- Enable the plugin in Krita - Preferences - Python Plugins Manager and restart
- Show the plugin docker: Settings - Dockers - AI Image Generation
- “In the plugin docker, click "Configure" to start server installation. Requires 10+ GB free disk space”
The models will be downloaded in ~/Library/Application Support/krita/ai_diffusion/
Last edited:
New models from civitai can be easily installed as well.
Open the AI plugin settings (small cogwheels icon docker/panel)
Create a new preset (+ button)
click on the folder button for AI model and the AI model location folder is opened.
download a new AI model from https://civitai.com/
place it in that folder and click the refresh button of the list of AI models. Then select it.
Optional LoRAs are installed the same way. Sliders control the strength of each LoRA that is installed in a preset.
Have fun!
Open the AI plugin settings (small cogwheels icon docker/panel)
Create a new preset (+ button)
click on the folder button for AI model and the AI model location folder is opened.
download a new AI model from https://civitai.com/
place it in that folder and click the refresh button of the list of AI models. Then select it.
Optional LoRAs are installed the same way. Sliders control the strength of each LoRA that is installed in a preset.
Have fun!
Unfortunately, there might be a serious issue with some resources used
huggingface https://github.com/Acly/krita-ai-diffusion/blob/main/ai_diffusion/resources.py
"Hugging Face, the GitHub of AI, hosted code that backdoored user devices" https://arstechnica.com/security/20...-ai-hosted-code-that-backdoored-user-devices/
huggingface https://github.com/Acly/krita-ai-diffusion/blob/main/ai_diffusion/resources.py
"Hugging Face, the GitHub of AI, hosted code that backdoored user devices" https://arstechnica.com/security/20...-ai-hosted-code-that-backdoored-user-devices/
No, there are no issues in this particular case, because all resources used by Krita's generative AI are safetensors from trusted sources. Safetensors are also a much safer format. You can look in the code itself that you linked to: but for the scale up ones all that are relevant are in safetensor format and from trusted sources.
It is indeed possible to hijack a safetensor file, but that hijacked version would then have to be distributed from another account/location.
This is comparable to downloading an executable from an untrusted website. No-one with an ounce of smart sense would ever do that without questioning the source.
And don't forget that Hugging Face hosts over half a million files. Like any other file hosting website we should never trust untrusted uploaders. Hugging Face also has anti-malware tools in place - which is more than what can be said for many commonly used online filesharing services.
Of course there will be people abusing sites like these. But stick to trusted sources and file formats that are (relatively) safe such as safetensors.
Also check this article, which shows that even the safetensor format can be compromised, but only a new file must be hosted on a different account or location.
hiddenlayer.com
In short: common sense prevails.
*edit* after reading some more there might be issues even with trusted sources. In this case I don't see issues, however. That said, if you are unsure: don't use it
It is indeed possible to hijack a safetensor file, but that hijacked version would then have to be distributed from another account/location.
This is comparable to downloading an executable from an untrusted website. No-one with an ounce of smart sense would ever do that without questioning the source.
And don't forget that Hugging Face hosts over half a million files. Like any other file hosting website we should never trust untrusted uploaders. Hugging Face also has anti-malware tools in place - which is more than what can be said for many commonly used online filesharing services.
Of course there will be people abusing sites like these. But stick to trusted sources and file formats that are (relatively) safe such as safetensors.
Also check this article, which shows that even the safetensor format can be compromised, but only a new file must be hosted on a different account or location.
Hijacking Safetensors Conversion on Hugging Face | HiddenLayer
In this blog, we show how an attacker could compromise the Hugging Face Safetensors conversion space and its associated service bot.
hiddenlayer.com
In short: common sense prevails.
*edit* after reading some more there might be issues even with trusted sources. In this case I don't see issues, however. That said, if you are unsure: don't use it
Last edited:
Tried Upscayl. Very slick. Thanks for the link, I'll be keeping this app installed. It does have a pretty noticeable color shift on images, but not to the point where a quick adjustment in Photoshop doesn't fix it. It has done wonders on low-res raster versions of logos clients have sent me.
Yes, I've tried them all. They all have varying degrees of color shift and fidelity.
They are all…fine I guess. But don’t expect magic. Topaz is good, but it (as many others) starts to struggle and introduce artifacts and inaccuracies when there are too many details
I forgot to mention here waifuExtension "The waifu2x on Mac" https://github.com/Vaida12345/waifuExtension
There is another thread about video upscaling https://forums.macrumors.com/threads/upscaling-1080p-footage-to-4k.2340662/
There is another thread about video upscaling https://forums.macrumors.com/threads/upscaling-1080p-footage-to-4k.2340662/
I have never paid $199 for a piece of software before Topaz. Version 2.4.1 is very fast on my M1Pro, and would not really be much faster on an M3 for the kind of single time processing I do. When I started printing my photos for gallery display at 13 x 19, I had needs to increase the size of the document, remove noise, and sharpen. Topaz does a terrific job. If the photo is bad to begin with, only generative AI would work, like Mid Journey, but Topaz is set against developing that and I agree with them. I don’t use them all by themselves. I usually start with Affinity Photo, move to Topaz, and do final work back in Affinity Photo.
genAI can be handy, but with upscaling it will "interpret" details. Here is a fun example: the Trevi Fountain in Rome. A 99 by 150px tiny image upscaled to 1584 by 2400 px! (done in Krita with the free AI plugin that upscales and "interprets").
Compare to a highres photo of the same fountain, and you'll see a lot of that "interpretation" going on. I probably gave it too much freedom here. ;-)
Sometimes the results are very good, though. But it's a bit of a crap shoot and also depends on the type of imagery and the AI model checkpoint used.
Compare to a highres photo of the same fountain, and you'll see a lot of that "interpretation" going on. I probably gave it too much freedom here. ;-)
Sometimes the results are very good, though. But it's a bit of a crap shoot and also depends on the type of imagery and the AI model checkpoint used.