kasothaphie/GenRecon ? reverse-engineered prompt

I want this repo turned into a working pipeline that can take a set of posed indoor photos, or frames from a phone video after camera poses are estimated, and reconstruct a detailed 3D scene mesh with textures. The main thing I care about is being able to point it at a scene folder, run one command, and get usable outputs in an output folder, ideally including a mesh export I can inspect or convert to glb.

Please set up whatever is needed to run inference with the released pretrained checkpoints, download those weights if they are missing, and make the setup as smooth as possible on a Linux machine with an NVIDIA GPU and CUDA. If there are multiple input modes, support the ScanNet style mode and a generic smartphone video workflow using COLMAP poses.

I do not need you to train new models unless it is absolutely required, but if training is already supported, leave it runnable and document it briefly. Add a clear README with exact steps, expected folder formats, and simple troubleshooting. Look up current docs online if you need to.