Understanding Video Resolution
A video consists of individual frames, and each frame is a collection of pixels. By multiplying the number of horizontal pixels (video width) with the number of vertical pixels (video height), we determine the video's resolution.
Typically, resolution is denoted as W x H, for example, 1920x1080 or 1280x720.
Generally, a higher video resolution implies more detail in your video. However, in the following section, we'll explore why a higher resolution doesn't always equate to better quality.
Exploring Video Resolution Upscaling
Upscaling video resolution has been around since the advent of digital videos.
What has evolved over time? Initially, upscaling was quite basic. Imagine a video with a resolution of just 2x2 pixels – a very tiny video indeed. :)
Consider a basic example where we use a simple upscaling algorithm to increase the resolution of a 2x2 video to 8x8, achieving a 4x upscaling factor!
With the upscaling to a larger resolution, we encounter a challenge. Many of the original pixels were duplicated to populate the new, larger pixel count. Although the resolution increased, not much new information was added.
In practice, more sophisticated upscaling algorithms are employed than the simplistic example provided. Bilinear and bicubic upscaling algorithms are among the most well-known. Now, let's explore how this scenario unfolds with a real video frame.
Consider a small video frame with a resolution of 200x170 pixels. We will upscale it by 4x to a resolution of 800x680 pixels using the bicubic upscaling algorithm.
After the upscaling, the resolution has significantly increased, but you might notice that the frame appears blurry and lacks sharp details. This is a limitation of the bicubic algorithm. While this algorithm is fast and can easily run on devices like mobile phones, it often results in less-than-ideal quality.
AI Upscaling
AI upscaling operates fundamentally differently, allowing for significantly more detail to be retained during upscaling.
Take the same frame upscaled by 4x, but this time using an AI upscale algorithm. Notice how the AI method recovers intricate details like freckles and individual hair strands. Unlike the previous method, the AI-upscaled version appears much sharper and less blurry.
AI Upscaling Filters
Understanding the benefits of AI upscaling over traditional methods, let's explore the upscaling options available in the Aimages app.
200% Upscale
This AI filter doubles both the width and height of your video. For instance:
- 1280x720 px becomes 2560x1440 px.
This filter is quicker and more cost-effective than the 400% upscale. It's ideal for videos already larger than 1280x720 px. For example, a video with a resolution of 800x600 px would be upscaled to 1600x1200 px using the 200% filter.
400% Upscale
This option quadruples both the width and height of your video. For example:
- 480x270 px becomes 1920x1080 px.
The 400% upscale is most effective for videos with resolutions lower than approximately 500x500 px. It's important to note that this filter is slower than the 200% upscale, as it has to generate a significantly higher number of new pixels.
Why don’t we always apply 400% upscale?
The decision between using a 200% versus a 400% upscale in AI video enhancement hinges on several factors:
Speed and Cost
While it might seem that more pixels automatically equate to better quality, the reality is more nuanced. The 400% upscale is significantly slower and more expensive than the 200% upscale. This difference is crucial, especially when processing large volumes of video or working under time constraints.
Diminishing Returns
There's a point of diminishing returns when it comes to upscaling. As you increase the upscaling factor, the AI has to work harder to recreate new details. Research has shown that upscaling factors above 4x don’t necessarily yield better results but do substantially slow down the process.
For instance, applying a 400% upscale to a Full HD video (1920x1080px) will produce a 7680x4320px video. However, this doesn't mean it will contain more meaningful details compared to a video upscaled by a lower percentage.
Light vs Pro Filter Versions
The AI Upscale filter comes in two versions:
- Pro
- Light
The differences between these lie in processing speed, cost, and the output quality.
The Light version is faster and less expensive but may not create as many details as the Pro version. The Pro version is preferable for extracting maximum details, especially from videos of lower original quality.
If the original video is already of high quality and you're seeking some extra detail while increasing the resolution, the Light version should suffice.
We recommend starting with the Light variant for quicker results at a lower cost in terms of Credits.
Ongoing Improvements
Video upscaling is a rapidly advancing field. We're continuously working on developing better algorithms that can produce more detailed outputs from the same input videos. However, such improvements require extensive research and development, so advancements may not be immediate.
In summary, choosing the right upscaling factor and filter version depends on the original video quality, the desired level of detail, processing time, and cost considerations. Always start with the Light variant for a balance of quality, speed, and affordability.