Frédéric Devernay

I had bugs with imageio/freemage when writing 32-bit FP EXR on Linux, and I'm not the only one, see https://github.com/imageio/imageio/issues/517 here's an example where I wrote the same data as...

This comment https://github.com/imageio/imageio/issues/517#issuecomment-657266692 also suggests using https://github.com/tvogels/pyexr

Update: - the bug is indeed on reading images with PIZ compression (either half or float - flags=imageio.plugins.freeimage.IO_FLAGS.EXR_FLOAT), and it happens only on Linux (macOS tested by me, WIndows tested...

Here's a test script: [imageio_issues_517.py.txt](https://github.com/imageio/imageio/files/8757050/imageio_issues_517.py.txt) Output on macOS: ``` in 1.0 1.0 test_write_imageio_float_piz_read_imageio 1.0 1.0 test_write_imageio_half_piz_read_imageio 1.0 1.0 test_write_imageio_float_none_read_imageio 1.0 1.0 test_write_imageio_half_none_read_imageio 1.0 1.0 test_write_pyexr_read_imageio 1.0 1.0 test_write_imageio_float_piz_read_pyexr 1.0 1.0...

tinyexr had a very similar issue: https://github.com/syoyo/tinyexr/issues/160

Same here. [AppPolice](https://github.com/AppPolice/AppPolice) (free and available from homebrew) also does the job of reducing the CPU usage.

Yes, this is a known issue, feel free to contribute a better est validation that using `diff` on the text output

sanity check: vanilla-nerf trains fine

Maybe we should aply those changes to the benchmarking scripts, or (better option IMHO) use a real-life dataset for benchmarking (T&T?).

For reference, here are two pytorch implementations of linear_to_srgb: - [NeRF-Factory](https://github.com/kakaobrain/NeRF-Factory/blob/f61bb8744a5cb4820a4d968fb3bfbed777550f4a/src/model/refnerf/helper.py#L25) (based on the multinerf code) - [Kornia](https://github.com/kornia/kornia/blob/406f03aa3da709dddeee4a6992a1158d2054fc63/kornia/color/rgb.py#L205)