Quantum Radiance Fields: A Quantum-Powered Photorealistic Rendering

Paper | Project Page | Vedio | Code
Feel free to use the chatbot at the bottom right of the webpage to inquire about any details you’d like to know reqarding this works.

NeRF vs QRF

• NeRF: given a 3D position (x, y, z), viewing direction (θ, ϕ), NeRF produces static and transient colors (r, g, b) and transparency values (σ).
• QRF: architecture replaces the same task with encoding circuit, parameterized quantum circuits, and quantum activation.

QRF Architecture

• Quantum Radiance Fields (QRF) with encoding circuits and quantum circuits produces colors (r, g, b) and transparency values (σ) given a 3D position (x, y, z) and viewing direction (θ, ϕ). Similar to the NeRF architecture, QRF enforces that the predicted σ is independent of view direction. Note that this schematic is a simplified quantum circuit with only 4 rotation gates around the z axis.

Optimization

• We supervise our QRF model by gradient descent using a simple 𝐿2 loss, and use Adam optimizer with a learning rate parameter α=0.001 to keep track of gradient moments over time to redirect the optimization trajectory.

Comparison

• Volume rendering of Chair, Ship, and Hotdog with training 100k iterations. Compared with NeRF, QRF has faster convergence, higher rendering efficiency, and higher rendering quality under the quantum integration.

Results on real-world captured data

• Camera position from COLMAP

Pine Tree in front of my home

• Training: 4k iterations
• The left one is inferenced by NeRF. The right one is inferenced by QRF.

Succulent Plan Potted

• Training: 4k iterations with 16 samples.
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  

• 3D mesh result with 64 samples.
  

Succulent_Plan_Potted by yfyangd on Sketchfab

Bench Park

• Training: 120k iterations with 64 samples.
  

• 3D mesh result with 64 samples.
  

Bench at Park by yfyangd on Sketchfab

Eluanbi Lighthouse

Eluanbi Lighthouse Model by yfyangd on Sketchfab