In addition to being a colorist himself, Dado Velentic has been working to develop tools to help other colorists (and really, all filmmakers) manage their finishing workflows. And at a recent talk at BSC Expo in the United Kingdom, Velentic has gone even further, sharing his current work that looks to bring AI and Machine Learning into the color suite.

Watch Velentic's entire talk below or read on for our takeaways.

One of the first things Velentic had to acknowledge was that most "auto-color" functions weren't very good. They rarely deliver on what they're designed to do: create an actual, proper match. However, there are algorithms that do a great job. Take "YouTube suggestions," for example, where the algorithm suggests a video for the user to watch next. YouTube can find eight videos out of trillions and recommend them to you. 

Screen_shot_2018-02-22_at_11Credit: Colorlab

Realizing this, Velentic's team talked YouTube into sharing access to the algorithm, and they began building Joi, a color assistant. They built thousands of bots, giving them a wide variety of approaches to color grading, and then built another bot to evaluate the results of each creation. They kept the top 2% of bots from every round of testing and recycled the rest. 

The team's conclusion was that it is, in fact, possible to train bots to help with grading. They might not be able to create looks, of course, but in terms of matching one shot to another to help with workflow, the robots do a great job. Assigning a robot to be lead colorist will be trickier, since there aren't any ways to spend a few years as an assistant if a robot carries that task out more efficiently. 

Screen_shot_2018-02-22_at_12The learning curve of the bots, being constantly corrected by humans guiding the process.Credit: Colorlab

AI is coming for jobs currently occupied by humans, and every industry is going to have to reckon with it. The film industry will not be immune. Velentic hopes to be back at BSC in 2019 showing off a working version of ColorLab Joi, built using the results of this machine-learning research.

For more info check out