Found this github.com/NoizAI/HelixWorld The README recommends an 80 GB GPU. A single-GPU BF16 run uses about 70 GB of VRAM. - Tested hardware: the real-time numbers (0.77 real-time factor at 768×512, 24 FPS) were measured on one H800 80 GB. On an A100 or H100 80 GB it should run about the same. - Text encoder: it uses Gemma-3-12B, which accounts for part of that memory. - Other requirements: Linux, Python 3.11, CUDA 12.x, and ffmpeg/ffprobe installed on the system. So the bottleneck is VRAM. Probably not running on consumer GPU machine anytime soon. But following just in case somebody implements a low VRAM option.