Single-Image 3D Scene Generator
Budget: $30 – $250 USD
I am in the final stretch of a computer-vision project that converts a single 2-D photograph into a fully navigable 3-D scene. About 89 % of the pipeline—dataset prep, training loop, and post-processing—is already finished. The missing piece is the generative model that actually predicts depth, completes occluded geometry, and exports the result in a standard 3-D file.
You will be working on an NVIDIA A6000 cloud instance that I have ready to go, so you can assume plenty of VRAM for large-scale diffusion, NeRF, or implicit surface models. I am framework-agnostic: if your best solution is PyTorch, TensorFlow, or Keras that’s fine, as long as the code is clean and reproducible.
Key goals
• Accept a single RGB image as input
• Infer depth and surface normals, hallucinate hidden geometry, and rebuild a watertight mesh
• Texture the mesh using the original image plus any learned in-painting
• Output a common interchange format (OBJ, FBX, or glTF—pick whichever integrates most easily)
Acceptance criteria
1. A Python script or notebook that runs end-to-end on the A6000 instance.
2. Trained weights or clear training instructions so I can reproduce results.
3. An example image and its generated scene demonstrating free-orbit camera movement with minimal artifacts.
If you have prior work with NeRF, Instant-NGP, or diffusion-based 3-D synthesis, please mention it—speed and quality are both important here. Let me know any additional resources you might need and your estimated turnaround time so we can wrap this project up.
You will be working on an NVIDIA A6000 cloud instance that I have ready to go, so you can assume plenty of VRAM for large-scale diffusion, NeRF, or implicit surface models. I am framework-agnostic: if your best solution is PyTorch, TensorFlow, or Keras that’s fine, as long as the code is clean and reproducible.
Key goals
• Accept a single RGB image as input
• Infer depth and surface normals, hallucinate hidden geometry, and rebuild a watertight mesh
• Texture the mesh using the original image plus any learned in-painting
• Output a common interchange format (OBJ, FBX, or glTF—pick whichever integrates most easily)
Acceptance criteria
1. A Python script or notebook that runs end-to-end on the A6000 instance.
2. Trained weights or clear training instructions so I can reproduce results.
3. An example image and its generated scene demonstrating free-orbit camera movement with minimal artifacts.
If you have prior work with NeRF, Instant-NGP, or diffusion-based 3-D synthesis, please mention it—speed and quality are both important here. Let me know any additional resources you might need and your estimated turnaround time so we can wrap this project up.