- Using 3D AI-generated assets can be incredibly efficient for game developers.
- However, these models can lead to high VRAM use, unnecessary CPU draw calls, and more when left unoptimized.
- Developers relying on AI for this purpose must ensure careful optimizations during each stage to avoid this problem.
It’s no secret that AI has now become closely tied to game development, with many studios using generative AI across different stages of game development.
The fact is that AI can even generate entire 3D assets, something that can be incredibly valuable to developers. However, these assets can become expensive when oversized textures, excessive materials, etc. reach production without proper reviews.
This is because generation and optimization fall under different stages of game development, so a convincing asset still needs a performance budget.
Why it matters: Game development has become ridiculously expensive and time-consuming as of late, so studios are looking at any possible solution to cut costs and timelines.
Fast Asset Creation Leads To A Different Bottleneck
Traditional asset production requires time that can be easily cut down by AI. For instance, using Meshy AI can help developers establish an early model through a simple description or reference image.
However, this also means that teams must carefully inspect which models should move forward. This requires inspecting geometry, adjusting textures, and creating levels of detail. If every acceptable result enters production, this can become a real problem later during development.
VRAM, Draw Calls, And CPU Costs
It’s no secret that modern hardware doesn’t come with the most amount of VRAM, even with the latest Nvidia graphics card generation.
Textures and mesh buffers consume VRAM, and a single asset can include base color, normal, roughness, metallic, ambient occlusion, etc. When combined, VRAM usage can go up very quickly.
Material slots also impact draw calls, with each material introducing extra rendering passes. This can further increase GPU overhead and CPU costs. Similarly, collision can impact CPU and system memory, especially when many objects interact with each other.
Even when average FPS looks acceptable, a large number of AI-generated assets can lead to frame-time spikes once VRAM becomes saturated. Therefore, all of these components must be carefully optimized.
VRAM Use Isn’t All About Polygon Count
Texture can consume a lot of VRAM, especially when high-resolution maps are not used. One 4K image may initially seem reasonable, but a scene with multiple props can quickly make for high memory demands despite the use of compression and texture streaming.
Here, developers must inspect imported textures instead of relying on source files. An object further away is unlikely to benefit from the same texture resolution as one close to the player, so developers must decide which objects to prioritize.
Material Slots and Draw Calls
Having too many material slots can make an asset incredibly expensive to render. This is because each additional slot could require another draw call.
As such, even when a single prop isn’t problematic on its own, having it repeated multiple times can cause issues later. This is why material consolidation is an important part of optimization.
Several surfaces can share one texture set or use a trim sheet, atlas, mask, etc. instead of separate materials. Keep in mind that draw calls also increase CPU overhead, so they require careful examination.
LODs Should Match The Asset’s Roles
Level-of-detail models are useful since they allow an engine to display a simplified version of an object when it occupies less space. This comes in handy when an object disappears quickly.
On the other hand, a larger object can benefit from a larger LOD chain for more detail. This means that an AI-generated asset should not immediately be pushed into production with its most detailed mesh displayed at every distance.
Developers should test LOD transitions during normal gameplay, in addition to reviewing culling distances.
Generate With Target Platform In Mind
A game like GTA 6 shows just how far developers can push modern hardware when optimizing properly.
This is precisely why optimizations should take priority at all stages of development, and when using an AI 3D generator, it’s a good idea to describe an asset’s intended role and target platform.
A prop for a mobile game and a PC title should be different. Terms like “simple silhouette,” “limited surface detail,” and “stylized low-poly prop” can produce tailored results through specific descriptions.
Of course, further inspection is still needed, but these descriptions can reduce the chance of beginning with an unnecessarily complex direction.
Collisions Can Also Hurt Performance
Using a complex render mesh can also lead to unnecessary physics cost. This is relevant when objects are put through collision checks.
Simple boxes or capsules are often sufficient here, and the appropriate solution depends on how the object is involved in gameplay interaction. Collisions should also match player expectations, so reviewing collisions with visualization tools enabled in the engine is important.
Ultimately, it’s important to create a test scene that contains the expected number of assets, lights, effects, etc. to measure VRAM use, frame-time, and other aspects that impact performance.
It may also be useful to test lower-end hardware, since performance becomes more relevant on these machines.
Thank you! Please share your positive feedback. 🔋
How could we improve this post? Please Help us. 😔
[Senior News Reporter]
Avinash is currently pursuing a Business degree in Australia. For more than 5 years, he has been working as a gaming journalist, utilizing his writing skills and love for gaming to report on the latest updates in the industry. Avinash loves to play action games like Devil May Cry and has also been mentioned on highly regarded websites, such as IGN, GamesRadar, GameRant, Dualshockers, CBR, and Gamespot.
Join Our Community
