andyyankai / psdr-jit

Unbiased Physically Based Differentiable Renderer

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

psdr-jit

Path-space differentiable renderer with drjit as the numerical backend. Old repo: https://github.com/uci-rendering/psdr-cuda

Installation

We use OptiX 7 to accelerate ray tracing. Make sure cudatoolkit (you may install it using conda) and an appropriate NVIDIA display driver that support OptiX 7 are installed. The latest version we tested was cudatoolkit==11.7. Then run

pip install psdr-jit

Local Intallation

To install psdr-jit locally, first clone the repository recursivly (to include the git submodules):

git clone --recursive https://github.com/andyyankai/psdr-jit.git

or

git clone https://github.com/andyyankai/psdr-jit.git
cd psdr-jit
git submodule update --init --recursive

Then compile and install psdr-jit via

pip install .

or

pip install --no-build-isolation -ve .

Notice that it would also install drjit as psdr-jit depends on it. cudatoolkit is also required.

Getting Started

import drjit
import psdr_jit

Example of Diff rendering

import cv2
import sys
import torch

import psdr_jit as psdr
import drjit
from drjit.cuda.ad import Float as FloatD, Matrix4f as Matrix4fD
from drjit.cuda import Float as FloatC, Matrix4f as Matrix4fC

sc = psdr.Scene()
sc.opts.spp = 32 # Interior Term
sc.opts.sppe = 32 # Primary Edge
sc.opts.sppse = 32 # Secondary Edge
sc.opts.height = 512
sc.opts.width = 512

integrator = psdr.PathTracer(3)	


sensor = psdr.PerspectiveCamera(60, 0.000001, 10000000.)
to_world = Matrix4fD([[1.,0.,0.,208.],
                     [0.,1.,0.,273.],
                     [0.,0.,1.,-800.],
                     [0.,0.,0.,1.],])
sensor.to_world = to_world
sc.add_Sensor(sensor)

sc.add_BSDF(psdr.DiffuseBSDF([0.0, 0.0, 0.0]), "light")
sc.add_BSDF(psdr.DiffuseBSDF(), "cat")
sc.add_BSDF(psdr.DiffuseBSDF([0.95, 0.95, 0.95]), "white")
sc.add_BSDF(psdr.DiffuseBSDF([0.20, 0.90, 0.20]), "green")
sc.add_BSDF(psdr.DiffuseBSDF([0.90, 0.20, 0.20]), "red")

sc.add_Mesh("./data/objects/cbox/cbox_luminaire.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,-0.5],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "light", psdr.AreaLight([20.0, 20.0, 8.0]))
sc.add_Mesh("./data/objects/cbox/cbox_smallbox.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "cat", None)
sc.add_Mesh("./data/objects/cbox/cbox_largebox.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "cat", None)
sc.add_Mesh("./data/objects/cbox/cbox_floor.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "white", None)
sc.add_Mesh("./data/objects/cbox/cbox_ceiling.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "white", None)
sc.add_Mesh("./data/objects/cbox/cbox_back.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "white", None)
sc.add_Mesh("./data/objects/cbox/cbox_greenwall.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "green", None)
sc.add_Mesh("./data/objects/cbox/cbox_redwall.obj", Matrix4fC([[1.,0.,0.,0.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.]]), "red", None)

P = FloatD(0.)
drjit.enable_grad(P)

sc.param_map["Mesh[0]"].set_transform(Matrix4fD([[1.,0.,0.,P*100.],[0.,1.,0.,0.],[0.,0.,1.,0.],[0.,0.,0.,1.],]))


sc.configure()
sc.configure([0])

img = integrator.renderD(sc, 0)
org_img = img.numpy().reshape((sc.opts.width, sc.opts.height, 3))
output = cv2.cvtColor(org_img, cv2.COLOR_RGB2BGR)
cv2.imwrite("psdr_jit_forward.exr", output)


drjit.set_grad(P, 1.0)
drjit.forward_to(img)
diff_img = drjit.grad(img)
diff_img = diff_img.numpy().reshape((sc.opts.width, sc.opts.height, 3))
output = cv2.cvtColor(diff_img, cv2.COLOR_RGB2BGR)
cv2.imwrite("psdr_jit_diff_debug.exr", output)

More Tutorial

gradient evaluation tutorial in tutorial folder

Object level optimization example: Neural-PBIR (ICCV 2023)

Scene level optimization example: psdr-room (SIGGRAPH ASIA 2023)

SDF+Nerf optimization example: to be released

Project based on psdr-jit

TBA

About

Unbiased Physically Based Differentiable Renderer


Languages

Language:C++ 53.9%Language:Jupyter Notebook 41.3%Language:C 2.6%Language:Cuda 1.9%Language:CMake 0.2%Language:Python 0.1%