ayaanzhaque/instruct-nerf2nerf

Leverages InstructPix2Pix diffusion models to iteratively edit pre-trained NeRF scenes via natural language prompts, optimizing the 3D representation with LPIPS and L1 losses in a Nerfstudio pipeline. Supports multi-GPU inference by isolating the diffusion model on separate devices, with memory-optimized variants (half-precision, no LPIPS) for constrained hardware. Integrates tightly with Nerfstudio's training infrastructure and provides configurable guidance scales to balance instruction adherence against image fidelity.

850 stars. No commits in the last 6 months.