Super high resolution images photo enhancing techniques used in TV shows and movies have often been criticized for not being realistic. However, enough research has been done in photo enhancing and the results are actually quite creepy. If you take a look at the latest AI photo upscaling technology developed by Google, you would be stunned by the results produced.
In a recent post published on Google AI Blog, the researchers in the company’s Brain Team have shared new breakthroughs they have made in image super-resolution. A machine learning model is trained for turning low-resolution photos into high-resolution, detailed photos. Two models were used for the purpose – Cascaded diffusion models (CDM) and Super-resolution (SR3).
What is Image Super-Resolution?
This model improves a low-resolution image using the Repeated Refinement process. The diffusion model takes an image of low-resolution as the input and builds a high-resolution image by picking up pure noise. To implement the entire process, the machine utilizes a process of image corruption where noise is added to a high-resolution image consistently up to a point where just noise remains. It then reverses the entire process for removing the noise until the target resolution is obtained in the image.
The results demonstrated by Google’s research team are absolutely impressive. It showcases how this technology can be used for improving the image quality of low-resolution images effectively. According to the post, super-resolution can have numerous applications including restoring old family portraits and making enhancements in the results obtained from existing medical imaging systems.
How to Download Super High Resolution Images From Google
One of the best ways of downloading super high resolution images from Google is simply using Google image search itself.
Search for images of your choice on Google image search. I will be using airplane as an example. Also, if you are on mobile, click the three dots on the top right hand side and check ” Request Desktop Site.”
Now that you are done changing the settings, click on the search tool on your new set up, a new tab with options such as color, size and more should pop up. Choose the settings of your choice.
If you have a high resolution of your choice, this setting should help you fix that before searching.
There are other search tools that can be used too. Another way is using Google Photos especially on desktop and mobile too.
When you want to edit photos using the Google photos, open the photo, click on the three dots and download the original.
How to get Super High Resolution Images for free
There are several free website you can download super high resolution images for free without the help of Google or going through all the settings as mentioned above on how to download super high resolution images from Google image search or Google photos.
Top ten websites to get free super high resolution Images free
How to create your own Super High Resolution Images For Yourself
There are numerous ways in which one can create its own high resolution images from Google which will be completely original, but the 5 steps on how to do this is best explained by Ian Norman from peta pixel
Steps on how to create the images by Ian Norman
Bring all the images you want to make high resolution into Photoshop as a stack of layers.
Use the “Nearest Neighbor” to resize your images to 200% height and 200% width.
Go on to auto align all the layers of photos
Set the layers to 1 per layer (First layer should be 1/1 100% opacity, second will be 1/2 so 50%, and so on.)
Sharpen your images by using the suitable amount of settings such as 200% for 4 image stack, 300% for 20 image stack, 400% for 10 image stack and more.
What is a Cascaded Diffusion Model (CDM)?
After achieving striking results from the Super-Resolution model, the Brain Team relied on the model of generating class-conditional images. As per the researchers, this model is trained on ImageNet data for generating natural, high-resolution images.
The team built CDM by cascading multiple diffusion models that are capable of generating images of increasing resolution. It begins with a standard diffusion model at the lowest resolution and is followed by a sequence of super-resolution models that can upscale the image in a successive manner. This process imparts higher resolution details to the image fed to the model. Along with SR3, the team also uses the data augmentation technique which is a new technique and also referred to as conditioning augmentation. This is said to be capable of making further improvements in the same quality results of CDM.
Effective results from AI-based models
The team used the CDM model to refine a low-resolution image of 64×64 to produce a 1024×1024 image. With the models showing promising results, Google is looking to make significant improvements to the natural image synthesis process. Though the technology does pose design challenges, it can have wide-ranging applications. As of now, the researchers are ecstatic about the results obtained with CDM and SR3 AI models. They have managed to push the performance of diffusion models to set new benchmarks in improving image quality.