So im looking into E-Cycles a paid version of optimized cycles as you know, for the purpose of speeding up renders. I dont own E-cycles but i do own Topaz AI gigapixel… it was about £100 but recently they switched to yearly subscription model which is something like £30 a year.
How can AI gigapixel speed up your renders and how could an AI upscaler conceivably be built into E-cycles just as Optix denoiser is built into E-cycles in a user friendly way?
Well i just rendered an example image… One of the blender demo files, using Cycles and the default setting of the file which is 200spp and the scene is using OIDN and some other compositor nodes such as dispersion, glare, gaussian blur, and adds a blurred noise to it, to add noise back in, so its not an ideal image for the AI denoiser to begin with.
Here is the render 1. at 1440x810. The render time was 2:03 -(on Ryzen 1700X plus GTX1070)
Here is the same image after being fed into the AI upscaler and then downsampled to 1080p with bicubic smoother. The upscale process took approx 10-20 seconds(but would be like greased lightning on an rtx card because of tensor core usage). The downscale is a software operation in photoshop that takes a fraction of a second.
Here is the render at full 1920x1080 cycles for comparison. and it took 3:40.
So i think this AI technique could be neatly built into something like E-cycles as a profile or something. Its a bit like DLSS in games.
I always upscale my renders 4X then downscale for the added sharpness and detail, its part of my post production workflow. but time consuming because you’re switching programs and such.
Here is another example this time using a different AI model within Gigapixel and this time using preserve details 2.0 on the downsample to 1080p.