Invoke AI has added two novel features to its AI-based image generation platform. According to the company, two new features—the Model Trainer and Control Layers—provide some of the most refined controls in image generation. Both apps provide users granular control over how AI develops and changes their photographs. Invoke also stated that it has achieved SOC 2 certification, which means that the company has completed multiple tests that demonstrate a high level of data security.
Invoke CEO Kent Keirsey spoke with GamesBeat, a local media outlet, regarding the platform's new features, which provide greater control and customisation over an image. The customised Model Trainer enables a company to train custom image generating models using as few as twelve pieces of its own content. According to Keirsey, this results in more consistent graphics that are congruent with a developer's IP, allowing the AI to create art with the same style and design features more frequently.
“We’re helping the models understand what we mean when we use a certain language,” stated Keirsey. “When we get specific and say we want this specific interpretation, what that means is we need anywhere from 10-20 images of this idea, this style we want to train… We’re saying, ‘Here’s our studio’s style with different subjects.’ You might do that for a general art style. You might do it for a certain intellectual property.”
According to Invoke, one of its goals is to provide increased security, which explains the SOC 2 compliance. Enhanced safety minimises the possibility that a developer's images will be exploited to help create another studio's intellectual property.
How to Train Your AI
Keirsey presented the second feature, Control Layers, which allows users to segment an image and assign prompts to certain sections. For example, a user can use the layer tool to paint the upper corner of an image and then instruct the AI to place a celestial body in that exact location. It enables creators to change the composition of their image and alter individual elements without impacting the whole image.
Each layer's cues can be refined and generated as any other AI image. However, the effects are limited to a specific part of the image. Control Layers also allows users to submit images to specific layers, and the creator can specify what elements of the image the AI should maintain - style, composition, colour, and so on.
Regarding how Invoke's new tools can be integrated into the game development workflow, Keirsey stated that most developers are cautious about the usage of AI, owing to copyright concerns.
“The human concept has to be there — a human sketch, a human initial idea. That will go to the point where you draw the line saying, ‘None of this is gonna go in the game yet. Until we can prove that we can get copyright, we’re not willing to risk it.’ The moment that you can get copyright, you’ll start to see that make its way into games… That’s why Invoke is trying to answer that for organizations, demonstrating human expression, giving them more ways to exhibit that, so that we can demonstrate copyright and accelerate that process,” Keirsey stated.