In this article, a developer and founder of the AI software explains the inner workings of the neural networks powering its software.Artificial intelligence has been transforming industries long before the recent burst in popularity of tools like Alexa, ChatGPT, and Stable Diffusion. Its core value lies in simplifying tasks, just like the convenience it brought to music lovers, for example. Think about the days of buying CDs, inserting them into players, and pressing various buttons on them just to listen to a song. Today, a simple voice command to a smart speaker analyzes your preferences and plays the perfect melody almost instantly. AI saves time and effort and enhances our daily lives.
The same is true for the professional field. Imagine a photographer spending 40 minutes on repetitive tasks like brushing, color correction, and minor blemish removal in Adobe Photoshop needed to retouch a single photo. This work “eats” the creative time they could’ve spent making photos. The question is, how can retouching time be reduced using available technologies?
When I was a photographer for nearly 10 years, I asked myself the same question after spending more time on post-processing than capturing images. Combining my technical knowledge and frustration with existing solutions, I started developing AI-based tools for streamlining my workflow. This journey led to the development of Retouch4me. Now we produce AI-based retouching software that takes the load of manual editing off the shoulders of photographers, retouchers, and videographers. Here is a view “under the hood” of the software.
Figure 1. Demonstration of Retouch4me Heal, Dodge&Burn, Portrait Volumes, and Skin Tone plugins’ functionality for photo retouching.The creation of the plugin starts with the formulation of the task for the neural network. It has to be specified in a certain way so that the algorithm understands it and can solve it.
We then collaborate with professional retouchers. We ask them to edit the image by completing a step in the retouching process. For example, removing eye vessels and providing us with the final retouched image. These images serve as the gold standard for the neural network to learn from and try to emulate.
Through training, the neural network internalizes the steps and techniques used by professional retouchers, enabling it to replicate their work and deliver professional-looking results. The work of the AI is based on a penalty system: the neural network strives to avoid unintended actions and gets rewarded for replicating the intended actions.
Figure 2. Demonstration of Retouch4me Heal and Dodge&Burn plugins’ functionality for photo retouching.
Each Retouch4me plugin represents a specific step in a retoucher’s workflow and is individually trained for that task (as mentioned earlier, removing skin blemishes or making teeth whiter). Our challenge lies in identifying the most effective training approach, as different retouchers may have unique methods. Choosing the right training data is crucial for the neural network to be trained correctly. This ensures the photo retouching software produces natural-looking results each time.
We used multiple frameworks for training neural networks previously. However, they did not fit our needs in the long run, either because of limitations or incompatibility with desktop use within Photoshop. This led us to develop our own custom engine. It granted us complete control and enabled the creation of innovative software pieces for photographers.
In a nutshell, the core technology behind Retouch4me plugins is the use of neural networks that emulate the work of skilled human retouchers. This approach saves time for the creatives and produces natural-looking results that avoid the artificialness often associated with AI.Article contributed by Oleg Sharonov is the creator of the neural network-based photo retouching software Retouch4me and 3D LUT Creator, and the founder of the IT company RELU OÜ.
Oleg is committed to contributing to the photography industry, simplifying the photo editing process with cutting-edge technologies. In his spare time, he enjoys taking photos, playing various musical instruments, and experimenting in the field of AI.