Hacker News Reader

Comment by ericjang | Hacker News Reader

Self-Plug: If you're interested in a Jupyter notebook tutorial on implementing a basic differentiable path tracer (a physically based rendering algorithm), I wrote up a tutorial here: https://blog.evjang.com/2019/11/jaxpt.html

View on HN

Great notebooks! While looking at the "cornell_box_pt_numpy.ipynb" one I noticed a few things odd:

- The input colours to the walls are scaled by somehow arbitrary values and they don't seem to match the spectral reflectance distributions converted to, for example, sRGB:

    White
    -----
    Jupyter Notebook             : [255, 239, 196] / 255
    sRGB (Scene-Referred Linear) : [ 0.73825082  0.73601591  0.73883266]
    sRGB (Output-Referred)       : [ 0.87468844  0.87351472  0.87499367]

    Green
    -----
    Jupyter Notebook             : [.117, .4125, .115] * 1.5
    sRGB (Scene-Referred Linear) : [ 0.06816358  0.39867395  0.08645648]
    sRGB (Output-Referred)       : [ 0.28953442  0.66418933  0.32540958]

    Red
    ---
    Jupyter Notebook             : [.611, .0555, .062] * 1.5
    sRGB (Scene-Referred Linear) : [ 0.49389675  0.04880702  0.05120209]
    sRGB (Output-Referred)       : [ 0.73132279  0.24476901  0.2508128 ]

The correct values, assuming a sRGB working space, would be the middle rows, i.e. sRGB (Scene-Referred Linear).

- The final image does not seem to be encoded for Output/Display, i.e. there is no gamma correction thus it is not displayed correctly.

Thanks, you are correct :) I chose the spectral reflectance distributions by tuning the values manually until it looked correct. And yes, I didn't do gamma correction (I deliberately ignored photometry related conversions) so that could explain why my tuning resulted in odd values.

Would be very easy to fix using the above numbers and adding proper encoding, even a naive gamma 2.2 would help :)

Here is a Gist for the spectral computations: https://gist.github.com/KelSolaar/fb1e09b56b54d37d90e1886aef...

(It uses Colour: https://github.com/colour-science/colour)

Always appreciate your posts. A couple projects in my lab are in Jax because of your MAML post.

Thanks, your comment made my day :)

At one point I was considering trying to write a differentiable renderer for 2d (cartoons) look. It was a bit too much work for my skill level, but I wonder if it got done already?

I mean something that would be able to output a vector representation of something.

Reply on news.ycombinator.com