The Tool That Sees the Hidden (Palantir)

About a year ago, while preparing a location for delivery, we faced stricter optimization requirements. At that time, the company lacked ready-made tools for Level Artists designed to solve this task. Interest in AI technologies was high back then, and I, being enthusiastic about new tech, had already tried using AI for writing code to solve simple problems. While implementing basic functions was possible, the user still had to carefully monitor the AI's actions, as agents didn't exist yet, and code had to be written directly in the chat and assembled piece by piece. That's when I decided to create my own tool for location optimization. I chose an approach based on searching for materials throughout the Unity hierarchy. By structuring this data, I was able to obtain detailed statistics on the usage of specific materials, textures, and prefabs in the scene.

This allowed for quickly finding objects used only once or textures without configured compression. These might seem like simple things, but even with such a tool, we stopped wasting time on blindly searching for problems. The screenshot shows the first version of the tool with minimal functionality. It was a very "buggy" prototype, but it could already perform its main task.

After that, the tool completed its task and was set aside until we finished another location. When the optimization question arose again, I returned to its development and implemented a new version in about 2-4 days. It gained new functionality, numerous filters, and duplicate search.

The updated version of the tool now showed the location's state, its weaknesses and strengths, much better. At times, this allowed optimizing up to 30% of the original materials and textures on the map simply by replacing them with frequently used analogs. This reduced the load with almost no impact on visual quality. One of the convenient tools at that time was the prefab mode, which allowed viewing textures used by a specific prefab. Also, the duplicate search mode helped eliminate similar materials with the same set of textures but slightly different settings—at times, we could have "50 shades of gray concrete."

And finally, after a year of iterative development, I can say the tool is, in a sense, ready. It was optimized for working with large scenes: the speed of calculating locations and processing a huge number of materials was increased, the analysis algorithm was completely reworked, the interface was improved, and much more. Listing all the improvements would probably take too much time.

But besides that, the tool acquired completely new functionality that can analyze the scene in volumes, show the load of zones by materials and prefabs, and build a heightmap of load based on the data. What does this give? Based on the new data after analysis, one can try to improve the framerate by relying on the tool's heightmap. It shows the most loaded blocks after analysis, as well as critical spots where there is a large spread in memory consumption between blocks.If anyone is interested in this tool and how it works, in the following posts, I'll try to record a video and show everything clearly.