Experimenting with low-fi asset gen in Unbound Loop. Mostly focusing on semantic parts splitting done at generation time, because 3D gen segmentation sucks. Would anyone be interested in giving this a try? 🤔

We're starting to take calls with solo professionals/studios that want to adopt a whole new way of generating 3D assets in real-time. Go to loop.unbound.io Click the feedback button and book a 30 min call directly with the team that'll be cooking for you 🧑‍🍳

In this video I'm asking an agent to generate a sci-fi crane in Unbound Loop. I then augment it with real-time AI, gen3D with Tripo and finally drop it in threejs. Lately I've been marveling at various solutions like img2threejs that use an image to instruct an agent, so I wanted to use our MCP connection in real world scenarios. I kept wondering if we should try to push our agent descriptors to increase details and spawn more primitives. But while testing this I realized we might have landed on a nice sweet spot already: Our input complexity from the agent-created SDF primitives is definitely low, which is actually ideal for manually tweaking things at the blocking stage. The reason why this works well with Loop's unique workflow is the output gain. Low input complexity does not mean low output detail. All I need to do after changing the agent proposed shapes is tweaking the AI strength and shape strength sliders. On top of the result from the real-time AI, I can continue refining with nano-banana without leaving the app. And once I'm happy with the image, I can move to 3D gen using tripo's smart mesh. At this point I can either modify the result again using SDF, or decide to take this out to my threejs game. I think we landed on a cool use case for agentic workflows, but curious to hear what other folks currently experimenting with this stuff have to say!