mshumer/gpt-prompt-engineer - GitHub vs GET3D | Nvidia

mshumer/gpt-prompt-engineer is an open source prompt engineering toolkit that generates, tests, and ranks candidate prompts for a task you define. You describe the use case, supply test cases, and the notebooks create multiple prompt variants, run them against every test case, and sort results with an ELO rating system starting at 1200. It ships as Jupyter notebooks you can run in Google Colab or locally.

Most prompt tools help you write one prompt at a time. gpt-prompt-engineer treats prompt selection like a tournament: dozens of candidates compete on your test cases, and the highest ELO scores surface the winners. Separate notebooks cover classification tasks, Claude 3 Opus with auto-generated test cases, and Opus-to-Haiku conversion for cheaper inference. Optional Weights & Biases and Portkey logging trace each run.

ML engineers, prompt engineers, and AI developers use it when they need reproducible prompt tuning instead of manual trial and error. The repo has 9.7k GitHub stars and runs on your own OpenAI or Anthropic API keys. It is free under the MIT license.

GET3D generates textured 3D mesh assets you can drop straight into a rendering engine, trained only from collections of 2D images. NVIDIA researchers built it to output explicit meshes with geometry and texture, not neural radiance fields that need custom renderers to use.

Most 3D generative models at the time either skipped textures, locked you into fixed topology, or required neural rendering pipelines to view results. GET3D uses differentiable marching tetrahedra (DMTet) plus adversarial losses on rasterized RGB images and silhouettes, so the output is a standard textured mesh. The trade-off is research-grade access: this is an NVIDIA Toronto AI Lab project from NeurIPS 2022, not a hosted SaaS with an upload button.

The model handles cars, chairs, animals, motorbikes, human characters, and buildings. It also supports text-guided shape generation via CLIP-based finetuning, similar to StyleGAN-NADA, and can disentangle geometry from texture through separate latent codes.