taichi_3d_gaussian_splatting icon indicating copy to clipboard operation
taichi_3d_gaussian_splatting copied to clipboard

working with 32bit or 16 bit floating point images for ACES color managed workflow.

Open luuude opened this issue 1 year ago • 2 comments

As far as I can understand this implementation uses only 8bit images as input. Would it be possible to take linear EXR or similar formats to preserve more data from the camera sensor?

luuude avatar Oct 13 '23 21:10 luuude

Indeed, you can definitely use linear EXR or similar high-bit-depth formats as input. While the current implementation predominantly handles colors normalized into 32-bit floats, modifying the input transformation handled by PyTorch should be straightforward. However, you might need to tweak certain parameters, particularly during rasterization. For instance, values like 1./255., set based on pixel sensitivity, might require lowering to accommodate the higher fidelity data.

wanmeihuali avatar Oct 17 '23 16:10 wanmeihuali

That is great news! Being a beginner in pytorch and python could you point in the direction of where to look to modiify the input transformation?

luuude avatar Oct 19 '23 12:10 luuude