Today’s generative AI systems can create some pretty remarkable images. But how do they fare with making HDRI Maps?
These specialized images are used in fields ranging from automotive rendering to architectural design. As experts in both generative artificial intelligence and HDRI Maps, with 20 years of experience in the latter, we decided to shed some light (pun intended!) on this subject.
The Two Parts of HDRi Maps
Fundamentally, HDRI Maps consist of two parts: the visual aspect of the image and the associated lighting data.
The visuals typically depict a real-world place, and they are often displayed as an equirectangular panorama. This method transforms a 360° image into a standard two-dimensional image.
3-D designers take these equirectangular images and load them into 3-D rendering software, creating an immersive environment where they can place digital twins or 3-D models.
The other significant component of an HDRI Map is the embedded lighting data, which can be used for things like image-based lighting.
This allows light from a real-world location to illuminate a 3-D model. Sometimes, designers use visuals from one source and use the HDRI Map solely for its lighting data.
What Can AI Create Today?
To see how well today's generative AI systems can recreate an HDRI Map, we decided to run an experiment using Midjourney.
Midjourney is one of the most popular, easily accessible AI image creators. We asked the system to generate an HDRI Map of a coastal road. We aimed to assess its capability in creating both the visual and lighting aspects of the map.
Here are the results. At first, the system made a 2D image--more like a backplate. To its credit, Midjourney did try to add in a mirror ball!
When we clarified that we needed an equirectangular image, Midjourney delivered one that actually looks decent visually.
Yes, it doesn't wrap properly for use in a real DCC. But it does gets aspects of the scene and the projection correct.
However, for the embedded lighting data, things were less impressive.
One of the most significant advantages of a professional HDRI Map, like those created by expert photographers, is its resolution and bit depth to display detailed lighting data.
In contrast, the image produced by the AI was of a lower resolution, at 1024 pixels square. This resolution is insufficient to include high-quality lighting data necessary for professional rendering.
Moreover, the generated image lacked the bit depth required for detailed lighting data. It best resembled a tone-mapped version of a traditional HDRI Map.
While the system's output is impressive in creating a visual of a nonexistent place, AI is not yet capable of producing a genuine HDRI Map suitable for 3-D rendering.
The Future of Generative AI and HDRI Maps
Although current image generators can't yet create HDRI Maps using AI, the necessary technology exists.
It's crucial to understand that present-day AI image generators are primarily trained on vast numbers of traditional photographic images.
They lack specialized knowledge about HDRI Maps essential for producing these assets with the required accuracy and detail.
However, if a custom image generator were trained on a large library of HDRI Maps, it could potentially create usable HDRI Maps with AI.
Another consideration for professional designers is that most current systems are trained mainly on images intended for social media or online use. These images often have dramatic lighting, and deeply saturated colors, and appear as if they've already undergone post-processing.
In contrast, professional 3-D designers generally need flatter images, giving them flexibility in post-processing to realize their creative vision.
Training generative AI systems on a library of professionally shot HDRI Maps or backplates could produce more accurate images with the flat colors and lighting needed for professional rendering, instead of the dramatic "Instagram look" of images from today's generative AI systems.
It's exhilarating to witness what today's systems can achieve, but we're even more excited about their potential future contributions.
Join our newsletter to stay informed as we continue exploring the capabilities of generative AI in 3-D rendering, and specifically the ways these systems can create HDRi Maps.
