I was wondering if they’ve said anywhere if they’re going to add a detelecine filter to the deinterlacing models? Probably a ways down the road, but it would be great to be able to get a film-based source back to its original 24fps.
Do the Gaia models plan on getting updated soon? I’ve used them a lot in the past, but now the results they produce just look really bad.
I use VE AI to enhance CG-related media (some irl stuff included) and every AI model except for Gaia are able to produce much better results.
Can I make a stupid question?
What happens using Dione on a progressive source? Does it mess up results?
The Gaia models are indeed in a dire need for an update, but depending on the source Gaia HQ still yield amazing results when feeding it good quality RL/3D-type inputs.
Gaia CG in the other hand is fairly similar to madvr’s NGU AA, in a sense where it upscales cartoon/anime-type inputs while staying true the source. (No nonsense “enhancements” are applied which only ruin the overall quality more than anything.)
1 Like
Ahhhh , that works. Thanks Robert. It’s not intuitive at all !! Perfect
After like a month upscaling hundreds of blu-ray ripped movies + old torrent ones. This is how I sum up all of the models that will give you the best result for the source input.
*Artemis HQ v11 => for good quality movies, films with at least 4000kbps+ bitrate for 720p, 6000kbps+ bitrate for 1080p. I use this most of the time for present blu-ray sources from 2010 till now.
*Artemis MQ + LQ => halo effects with trees and bushes for videos under 720p only. Best for 720p + 1080p with medium compression + artifacts only.
*Artemis AA + Gaia CG => same thing with minor differences. Pick Artemis AA for fast + medium quality. Pick Gaia CG for slow + good quality. I only use these models when my source input is 480p or 720p (bad quality). This will slightly sharpen and reduce noise + block without having weird faces or blurry effect from Artemis MQ + LQ. I recommend using Artemis AA if movie, Gaia CG if anime.
*Gaia HQ => Artemis HQ on steroids, but motion consistency is what I don’t like. Recommended for 720p and above with medium compression and artifacts. Best with 1080p with <1% motion flicker. Good for slow movies, bad for fast paced movies. Only use if Artemis MQ + LQ can’t please you. Otherwise, use Artemis HQ if the input is already considered good.
*Theia Fidelity + Detail => Artemis with customization, only use if you know what you’re doing. That’s it. If you don’t, ignore it.
*Dione => sorry, still not have enough time with it since most of my sources are Progressive.
–TL;DR:
*720p/1080p above medium/good quality => Artemis HQ for fast, Gaia HQ for slow but more details (consider using it if your source has medium quality)
*480p/720p under medium quality => Gaia HQ
480p/720p @&$ quality => Gaia CG (slow speed-good quality-motion flickering)/Artemis AA (fast speed-medium quality-less motion flickering)
9 Likes
This sounds like a very promising workflow. Could you provide any additional details such as what tools you used for edge detection and overlaying? Thanks for any help!
Here’s an example script I used recently:
function GCTM(clip input) {
input.Levels(512, 1, 1023, 0, 1023, false)
one = last
input.Levels(0, 1, 512, 1023, 0, false)
two = last
Overlay(one, two, mode="add")
}
artemis = FFVideoSource("youngones_s02e04_artemis.mov").Crop(0,0,0,-2)
gaia = FFVideoSource("youngones_s02e04_gaia.mov").Crop(0,0,0,-2)
Subtract(artemis,Blur(artemis,1))
GCTM()
Levels(20,1,96,0,1023,false)
Overlay(last,Blur(1),mode="add")
Overlay(last,Blur(1),mode="lighten")
Greyscale()
mask = last
Overlay(gaia,artemis,mask=mask)
Spline36Resize(1440,1080)
Prefetch(16)
This is assuming ProRes renders so the calls to Levels() are in 10 bit.
1 Like
You’re a legend. I’ve bookmarked your reply for future reference. Thank you.
1 Like
非常不错的软件
I tried it, it actually makes the footage sharper, sometimes better than Artemis but you won’t get the full detailed/sharp result as you would get from an interlaced source (since it reads both fields and does a lot of AI calculation).
Has anyone found a good use case for the Theia models? Even at zero sharpness they are too sharp for my taste, resulting in unnatural sharpness. And I could not see any real difference between both.
Personally I think they should just be removed tbh. At least Gaia still has a reason to exist, but Theia doesn’t do anything that Gaia or Artemis can’t do better.
1 Like
Thank you very much, I was asking so because I noticed that Dione creates many interlacing artifacts with that source so I was wondering if it could possible to deinterlace with QTGMC and then upscale using Dione to keep details and have a sharp image, I tried Artemis a couple of times but it does a too strong denoise.
version 1.9.0 was an amazing improvement (speed-wise) over the prior versions, so I’m looking forward to working with this new update.
Cheers,
-Lou
Anime.
Hi,
I used Theia Fine Tune with just Denoise = 15 and sharpen = 50 (after some visually tests).
I tried Theia Fine Tune and it’s true that visually it’s quite complicated to prove that there is a big difference.
I am trying to create a metric for myself from a test pattern to find the best results between the different models with quantified results according to a source clip.
I’m thinking of using an ISO_12233-reschart test pattern, not knowing if it’s the best. (It’s a test pattern rather for cameras / APN => Lens). Knowing that it is a test pattern rather 8k, in 4K, there is already a small loss (a little blur and some aberations (frequencies?)).
If i place the chart on a TL 4K (Davinci Resolve) and render one or 2 seconds in DNXHR 444 12b, then ask VAIE to do an 8K with different models / settings, maybe visually after i can measure to what degree VAIE encodes to recover loss without creating oversharpening.
I’m looking … but not easy.
(The original tif weighs 743 859 KB, here it is a very compressed jpg which weighs 1230 KB)
I am not sure if a lens test chart will be of any help. Normally those are created in vectors and then exposed via an imagesetter in 5080 dpi. Those patterns are designed to break camera sensor imaging and lenses for evaluation but do not resemble any real world image for which most VEAI models are trained.
In my own test I have found when using the preview the result depends on the frame I start the review in a clip. With Artemis MQ for example its de-haloing differs quite significantly on the same clip if I start the review at frame x versus maybe 50 frames at frame y later (when the content has changed a bit.
2 Likes
Thanks. I do not use VEAI with cartoons/anime, so I had no idea it works well for that kind of video.
I dont like grain. So I made it 0 and save as default. But it wont work like in v1.9.
When I open VEA 2.0 again. Grain is still there. So save as default somehow ignore grain.
And sometime even if grain look zero, output would have grain effect except first video.
Somehow second and after videos grain settings is stays even it looks zero in gui.
Probably grain settings must be change and set zero for every video.
Please look at this.
2 Likes
