fmannan / diffrend

Differentiable Renderer

DiffRend: An Experimental Differentiable Renderer

Setup

python setup.py develop --user

Usage

Command line:

python diffrend/render.py --use [gl|np|tf|tch] --scene <scene-description-filename>

Using the TF based optimizer

from diffrend.tensorflow.renderer import render

# generate splats

splat = ...

# setup the scene here
scene = splat_to_scene(splat)

# render
res = render(scene)
image = res['image']
im_depth = res['depth']

# use
# ...

For other versions use:

from diffrend.numpy.renderer import render

Code Organization

Numpy, Tensorflow and PyTorch versions of the renderer are in

diffrend/numpy
diffrend/tensorflow
diffrend/torch

folders respectively.

model.py: For loading and transforming models

3D Model and Geometric Primitives

Geometric primitives:

• Plane
• Disk
• Triangle
• Sphere

Mesh files: Currently OBJ and OFF only for triangle meshes.

Lighting/Shading/Material

Scene Description File

Scene Description File Format

Rendering

Forward Pass

Gradient-based Optimization

Memory/Speed

In most cases there will be more pixels than geometric primitives. Pixels are rendered independently of each other. So if the GPU runs out of memory limit then tiled rendering might fix the memory issue.

Render Server

Render Server