Seemingly overnight, Artificial Intelligence tools are now absolutely everywhere, running all sorts of applications in most industries, even in places where you don’t really see them.
OpenAI’s ChatGPT, the text-based natural language model that responds in a human-like way to the prompts that it’s given, is perhaps the most well-known of this new breed of AI tools. The fact that millions of people are using ChatGPT worldwide every day attests to this.
However, using AI to create digital art is also a very popular application, thanks to services such as Midjourney and Stable Diffusion. In this blog, we’re going to dive into Stable Diffusion Upscale, so that you can stay ahead of the curve when it comes to collaborating with AI, and present your visions to the world exactly how you want them to be seen.
Stable Diffusion Upscale
AI has completely changed the playing field when it comes to living our daily lives. Over the past six months or so, people have been using different artificial intelligence models to assist them with completing their tasks with less effort.
And although alarm bells have been raised recently by tech giants like Tesla, Twitter, and Neurolink’s Elon Musk about the potential threat that AI poses to the continuity of the human race as a whole, there’s no arguing that this genie has truly been let out of the bottle — and it will never go back in.
Today, anyone with a computer and/or smartphone has the ability to interact with artificial intelligence like never before. And that includes those of us who are interested in art, because there are now AI services that are able to generate images from accurate text prompts and keywords.
This is why digital art is now becoming a winner with many budding and dedicated artists worldwide. Midjourney and DALL-E are probably the most well-known of these AI image generators, followed closely by Craiyon and Stable Diffusion.
But how well does Stable Diffusion actually work when it comes to producing quality results? And is that the reason why you need to worry about upscaling the images that it generates?
Let’s take a closer look at the upscale debate…
The What And Why Of Upscaling
Although it can be very useful, artificial intelligence isn’t exactly perfect — well, not yet anyway. And so, this is why many users of Stable Diffusion have noticed some issues with inaccuracies with the AI model’s output, as well as visual degradation of generated images in some cases.
For instance, Stable Diffusion already has a reputation for not being great at generating images that contain human limbs. It also has an issue with faces, which tend to be quite similar, due to the lack of varying ethnic features that were represented in the database used to build the AI.
Then there’s the fact that many users believe Stable Diffusion’s main problem could stem from the fact that the AI was trained with a dataset of images that were only 512 × 512 pixels in resolution.
This resolution was later increased to 768 × 768 with the Stable Diffusion 2.0 update. But even though the AI is now able to generate images at this higher resolution, the resulting output can still not appear exactly as you would like it to.
So, although the images the AI model generates may have the feel and tone that you were after, the most common complaint against Stable Diffusion is that the final size of the output is generally quite disappointing and very small, which makes it difficult to use for other digital purposes.
Although the images the AI model generates may have the feel and tone that you were after, the most common complaint against Stable Diffusion is that the final size of the output is generally quite disappointing and very small, which makes it difficult to use for other digital purposes.
And this is why upscaling your images after they’ve been generated by Stable Diffusion may be an important step that you need to take — especially if you’re a content creator or digital artist, and you want to display your digital art on your website or for other marketing purposes.
Size Does Matter
As we’ve just mentioned, by default, Stable Diffusion’s image size is now 768 x 768 pixels. But as any digital artist will tell you, this is actually quite small by today’s standards. Even most social media platforms require at least a 1080 x 1080 resolution for posting images, in order for them to appear acceptable when on screen.
By default, Stable Diffusion’s default image size is now 768 x 768 pixels. But as any digital artist will tell you, this is actually quite small by today’s standards. Even most social media platforms require at least a 1080 x 1080 resolution for posting images, in order for them to appear acceptable when on screen.
Lucky for us though, there are a few dedicated online tools available that you can use to help you upsize and upscale your Stable Diffusion-generated images at the click of a button. One such program is the popular TinyWow Image Upscaler.
Unlike Stable Diffusion, AI upscale tools like TinyWow have been trained with massive amounts of image datasets. This means that they work on ‘recovering’ the missing details in the images you upload.
So, to use TinyWow, all you need to do is upload or drag the Stable Diffusion image file from your computer or smartphone onto the platform, and then select Increase. After which, TinyWow will present you with the option to download your new upscaled image.
It’s as simple as that!
TinyWow, unlike some of the other online image upscale tools, is completely free to use. If you’re a digital artist, then this step alone could save you loads of time by eliminating any manual work that may have been required to fill in the gaps of your Stable Diffusion images when you enlarge them.
We decided to challenge TinyWow to a test to see how well it works. But first, here was the prompt we gave Stable Diffusion to create our desired image:
Imagine a tropical setting where avocados grow in abundance. The sun is shining in a deep blue sky. The colorful birds are singing. The air is filled with the sweet scent of frangipani and ripe avocados. Create an image that captures this idyllic essence. Bright lighting and tropical color palette would best convey the warmth and vibrancy.
We then visited TinyWow to upscale our Stable Diffusion generated image. The platform gives you the option of upgrading your image 2x or 4x. For this example, we chose the 2x option. Here are the two images, zoomed in, side by side:
As you can see in the left-hand image generated for us by Stable Diffusion, the pixelation can be seen once it has been zoomed in, while the TinyWow copy on the right-hand side has clearly been upscaled.
To give another example, we gave Stable Diffusion this prompt:
Picture a sunny spring day in a city park filled with cherry blossom trees. A French bulldog is sitting on a picnic blanket, looking up at the trees, watching as the petals fall gently to the ground. In the background, there are other people and dogs enjoying the beautiful day. Your task is to create an image that captures the joy and beauty of this moment. Make the French bulldog stand out in the image. Photorealistic.
And side by side, with the original Stable Diffusion image on the left, this is what TinyWow was able to produce after enlarging the image — this time with the 4x upscale enabled:
So, as you can clearly see, although Stable Diffusion’s AI image generator does a great job at creating stunning images from your prompts, if you need an enlarged version for any other purpose, then TinyWow is a great tool to supplement your digital artwork.
Being an open-source AI model though, someone somewhere will always be working to make Stable Diffusion better, like this beta Upscaler program. So, if you’re a bit of a tech head, then you can visit the Stable Diffusion Upscaler Demo to try it out for yourself.
In addition, Stable Diffusion 2.0, which is on open-source release, includes an Upscaler Diffusion service that is able to upgrade image resolutions by a factor of 4x, allegedly up to 2048 x 2048 and higher.
As Stable Diffusion’s researchers continue adding more datasets, and fine-tuning the AI model’s algorithms, then the generated image size will more than likely keep increasing, and the degradation issues will become less persistent.
But until then, you now know where to go for help with your Stable Diffusion upscale needs when the text-to-image prompts need more detail and enlarging.