Local AI Orchestrator - 3D Asset Edition
- 6 hours ago
- 1 min read
Updated: 5 hours ago
I have spent the last few years exploring AI/VFX hybrid workflows in ComfyUI, and today I am sharing my latest development.
I built a standalone desktop orchestrator that turns local generative models into a real asset pipeline. It drives models entirely on your own GPU. Write a list of props or upload your own concept art, and it returns a pack of textured 3D meshes, turntables, and contact sheets.
No manual FBX round trips: finished assets land in a running Houdini session as a built node graph, and (after some manual tweak) a companion script pushes your rigged, animated character from Houdini straight into your Unreal level — skeleton, animation and material, with the clip already playing on the actor.
It also fixes the part everyone gets stuck on: setup. It works around the Python 3.12 / 3.13 mismatch that breaks Trellis in ComfyUI by assembling its own 3.12 backend rather than touching the install you already have. It sizes the local prompt writer to your exact VRAM (8, 16 or 24 GB), and gives you a 3D viewer to judge topology without opening any other software. Zero telemetry, and machine-readable AI markers embedded in every image it writes.
Windows, NVIDIA, 8 GB VRAM minimum. No macOS or AMD yet.
I am opening a closed beta for cohort 1, picking for fit rather than speed.
Applications close 16 September 2026.