That looks indeed impressive. The last time a friend of mine showed me the results, it looked blurry in some areas and with heavy high frequency “texturing” in some other areas, giving the impression of looking through a bad lens. Especially on renders with depth of fields were it would suddenly add sharp details in middle of an out of focus area.
Your example looks nearly like it’s rendered at that resolution (apart from the cafe machine buttons and some blur on the metallic parts. Your picture is at 60% zoom, so more like 2k to 4k and has everything in focus, which may explain the very good results. Nonetheless, very impressive, I might have another try at scaling AI later as their are some good open source ones.
No the Seed option is not active. If you want I can send you the blendfile. But i don´t know how to send you a PM. I have no option … it could be because my account is new?
You can send me an email by answering your Gumroad emails about E-Cycles. I’ll have a look if I can optimize such case. I’m wondering were the slight flickering comes from if neither the camera nor the seed are animated, so it will be definitely interesting to investigate.
Ich habe die Datei runter-geladen, danke. Obwohl es eine no-reply Adresse anzeigt, die Email wird an mich weitergeleitet. Auf jeden Fall hat das bis jetzt für alle funktioniert. Ich würde einfach probieren auf “antworten” zu klicken und gucken an welcher Adresse es geht/gehen würde.
You are allowed to do it as a user, but for to distribute the Optix denoiser with E-Cycles, it should be open sourced with a GPL-compatible license first. But the addon from remington graphics should work with E-Cycles.
Has someone else tried to use AI Denoise with Volume Scatter? In my case, the denoiser always kills the fog. Do i have to add a seperate Volumepass in the rendersettings and combine it back in inside the compositor, or is there a better solution for this?
Level 1 and 2 may give better results with fog. With level 3, you can indeed activate the direct and indirect volume passes, then go to the add menu → filter → denoise. Combine the direct and indirect output of the volume pass with add menu → color → mix node in add mode. Feed the output of that add node to the denoise node and then add everything together. But volumes don’t have normals nor albedo, so the denoiser may overblur things. In this case, a mix node with the noisy volume image may help keep details (see second screenshot).
Both E-Cycles 2.8x and 2.79x now have the May update. The new features so far on top of the usual improvements made in master are:
Around 10% faster rendering for 2.8x and lower memory usage (ported from the 2.79x version which already had it since 3 months):
Up to 2x faster denoising on multi-gpu setups when using the old denoiser (AI denoiser already was faster in this case). This is new on both 2.79x and 2.8x.
The builds are available for Windows, Linux and Mac.
I already worked on GI caching, when it will be ready for release is yet to be seen. I’m first working on making adaptive sampling ready for June for now and making tiled OIDN to allow denoising huge images with limited memory.