Chaos Corona 13 for 3ds Max - Daily Builds Discussion

[TomG]

I can quote Alex's really good suggestion of running internally downresed input so the user can have the primary image be native high-res and AI enhancement up-resed.

I still miss the internal AA down-resing from Corona Alpha :- )

Oh the internal AA down resing, that was before my time - is it at all similar to what is shown in this recent video?

(basically, render to 2x resolution, then downscale to final image size). This is an idea I am considering putting on the cards for development, basically skipping the Photoshop step of downsizing, so am interested to hear thoughts.

[pokoy]

I tested this and yes the result is better than the original. AA of finer details and highlights and bitmap filtering are obviously sharper due to downsampling.

There are limitations though which aren't obvious:
- Complicated shader trees tend to resolve detail (bump, procedural details in noise maps etc) at a later stage in the render progress. With only a few passes the result will be less surface details.
- Some of Corona's sampling-based maps like AO and Curvature will also refine at later stages. I got some heavy differences between the two images where AO was much less pronounced and detailed in the downsampled image (with less passes). If you're using those kind of maps, it's best to still allow more samples.
- Refraction details also tend to take a bit longer, so you trade sharper details (implying more detail) for less actual detail and accuracy.

However on simple scenes and simpler shader trees the workflow results in surprisingly more detail when compared 1:1.

Interestingly, in the first days of Corona there were no image pixel filters and this was exactly the way Corona did it - by rendering twice the resolution internally and displaying the downscaled image in the VFB. EDIT - Oh wait, this was already mentioned above...

[lupaz]

Or just focus on the ideas with the most votes?
I know that the ai effort is Chaos-wide, but there are many other things on the list and these features are very much replaceable with other existing and better products out there.
I don't mean to tell you how to do your product. This is just my two cents.

Just a note on this that no-one on the Corona team spends any dev time at all on the AI itself. All we do is the UI and hooks to get an image sent up to Chaos Cloud (well, and created a very easy new render element, which is basically just three other render elements combined). So you are seeing a feature that is as close to "free" as it gets in terms of development time, so much so that not adding it wouldn't make any sense :)

Got it. Thanks for the note. Nice perk of the merger.

[Okmijun]

I can quote Alex's really good suggestion of running internally downresed input so the user can have the primary image be native high-res and AI enhancement up-resed.

I still miss the internal AA down-resing from Corona Alpha :- )

Oh the internal AA down resing, that was before my time - is it at all similar to what is shown in this recent video?

(basically, render to 2x resolution, then downscale to final image size). This is an idea I am considering putting on the cards for development, basically skipping the Photoshop step of downsizing, so am interested to hear thoughts.

I've learned this trick from Juraj, back in the early days of corona, and yes, I think this could be easily incorporated into daily builds.
some like "render 2/3x bigger and downscale it" option, for final output.
+1
So, basically corona will become 5-10x faster within one single click, with 2-5% loss of details, maybe not even closer.
My advice is HURRY UP, as you slightly loosing the speed battle against GPU renders...
With this little improvement/ hack you are in the game againa...

[lupaz]

I can quote Alex's really good suggestion of running internally downresed input so the user can have the primary image be native high-res and AI enhancement up-resed.

I still miss the internal AA down-resing from Corona Alpha :- )

Oh the internal AA down resing, that was before my time - is it at all similar to what is shown in this recent video?

(basically, render to 2x resolution, then downscale to final image size). This is an idea I am considering putting on the cards for development, basically skipping the Photoshop step of downsizing, so am interested to hear thoughts.

I always hated denoising but this sounds pretty amazing. It would be great to have it done under the hood with just one click...As an option I'm guessing.

EDIT: And to me this is a much proper use of AI for a Rendering engine than improving the look of 3D people. I think it would be best if Chaos has a separate application for that kind of work (the AI enhancer), and not integrated into a rendering engine.

[marche3d]

You know what would be awesome? Once the Corona AI Enhancer is fully polished, it would support image sequences too, so we could enhance people in animations as well.

[Kelevra]

I can quote Alex's really good suggestion of running internally downresed input so the user can have the primary image be native high-res and AI enhancement up-resed.

I still miss the internal AA down-resing from Corona Alpha :- )

Oh the internal AA down resing, that was before my time - is it at all similar to what is shown in this recent video?

(basically, render to 2x resolution, then downscale to final image size). This is an idea I am considering putting on the cards for development, basically skipping the Photoshop step of downsizing, so am interested to hear thoughts.

Yes, please.
Skipping the downscale process in Photoshop would be amazing.

[Sandertype]

+1 to having this auto rescale with the ai denoiser done under the hood.

[Mr.White]

+1 also from me. It works fantastic and the benefits are many (saved time especially). Would be great to see it implemented directly in Corona.

[alexyork]

Not to derail the topic about the AI thing, but that video about the intel denoiser used with tent really changed the game. We're finding significant detail boost at lower passes using this method, and that's without actually rendering at 2x scale and downscaling back down. It's meaning we can render animations natively at 4k in the same general time as we'd have rendered 1080p animations, because the image detail and denoising results with the Intel denoiser (with Tent) are so impressive. And of course the benefits are there too no matter what resolution you're working at, so we're also using this method now for general work e.g. 6k/10k renders. The whole 2x scaling thing is more faff than it's truly worth IMO. Also what's not discussed in that video AFAIK is RAM usage at higher resolutions - something to consider.

[romullus]

Not to derail the topic about the AI thing, but that video about the intel denoiser used with tent really changed the game.

May i ask, what filter and denoising combo you were using before? I always set Intel denoiser and tent filter as default after each new version install since it doesn't work with HQ filter and i find it quite amusing that apparently it's a new discovery. Oh, and i completely agree with you about that rage about 2X scale - i did some short tests and my findings was that out of 3 very different scenes it kinda worked ok'ish in one of them, but the other two were clearly inferior to just simple rendering at native rez.

[dj_buckley]

Not to derail the topic about the AI thing, but that video about the intel denoiser used with tent really changed the game.

May i ask, what filter and denoising combo you were using before? I always set Intel denoiser and tent filter as default after each new version install since it doesn't work with HQ filter and i find it quite amusing that apparently it's a new discovery. Oh, and i completely agree with you about that rage about 2X scale - i did some short tests and my findings was that out of 3 very different scenes it kinda worked ok'ish in one of them, but the other two were clearly inferior to just simple rendering at native rez.

Glad I wasn't going mad back in 2021 https://forum.corona-renderer.com/index.php?topic=32584.msg182506#msg182506 it was around this time I made my defaults Intel/Tent and never looked back

[alexyork]

Somehow it just missed us, this knowledge! Been using the default Corona HQ denoising + HQ filtering since... forever. One naturally assumes that anything native/Corona and "high quality" and which is also the default must be the best, for day-to-day use, and of course we've never had any issues with it, therefore no incentive/reason to try anything else. It's just this recent vid that caught our attention recently which caused us to change. But yes, the difference really is night-and-day. I wonder if the corona method simply needs to be improved if it's to remain the default... or just make Intel+Tent the default if better?

[dj_buckley]

Somehow it just missed us, this knowledge! Been using the default Corona HQ denoising + HQ filtering since... forever. One naturally assumes that anything native/Corona and "high quality" and which is also the default must be the best, for day-to-day use, and of course we've never had any issues with it, therefore no incentive/reason to try anything else. It's just this recent vid that caught our attention recently which caused us to change. But yes, the difference really is night-and-day. I wonder if the corona method simply needs to be improved if it's to remain the default... or just make Intel+Tent the default if better?

That's what I was repeatedly told when questioning it too, in that thread I posted you can see the other response were convinced with comments like "I rarely sharpen now with Corona defaults".  Intel + Tent is my default in my Maxstart file, i like the freedom of Tent too as you can adjust the Width(Px) to your liking, higher values softer, lower values crispier

[romullus]

To be honest, i prefer HQ filter over any other, too bad it's not compatible with AI denoisers and since having noise free image in a fraction of time is more important than having nicer looking noise, there's little sense to go with anything else but Intel denoiser + Tent. I agree that Corona team could evaluate an option to make it the new default.