It seems that pytorch3D only support the rasterized renderer. Will it support the ray-tracing method in the future version?
So technically mitsuba2 natively supports gradients for pytorch tensors. However there are many functions lacking (batched rendering, multi-gpu, differentiation w.r.t. vertices etc.) that make it not realistic to use for training, but these are supposedly on their roadmap. The documentation is also quite sparse, so your mileage may vary.
Most helpful comment
So technically mitsuba2 natively supports gradients for pytorch tensors. However there are many functions lacking (batched rendering, multi-gpu, differentiation w.r.t. vertices etc.) that make it not realistic to use for training, but these are supposedly on their roadmap. The documentation is also quite sparse, so your mileage may vary.