Fire3D reconstructs simulation-ready 3D scenes from images

Fire3D bridges the critical gap between visual perception and generative 3D modeling, turning what used to be a multi-hour pipeline of NeRF optimization and manual asset cleaning into a fast, one-minute feed-forward pass. By bypassing iterative per-scene optimization, it generates complete scenes and textures orders of magnitude faster than prior interactive reconstruction methods. Its object-compositional approach outputs decoupled, independently transformable meshes rather than monolithic radiance fields, making environments immediately usable in physics engines like Isaac Gym or Blender. The high-compression HC-VAE enables scalable parallel diffusion and mesh decoding across up to 16 entities per batch, though reliance on high-memory workstation GPUs (up to 96 GB VRAM) currently limits local deployment on consumer hardware.