Please replace Gaia CG. Something CUGAN-like would be ideal. Otherwise, please update the model. It’s totally useless.
Even if they updated it, they would leave the old version available for use. (Or non-use in your case.)
I have no idea what CUGAN is, but it sounds like your request should be: make a CUGAN-like model. Give Topaz details on what that should look like or be able to do. No need to drag Gaia into it.
A request to update Gaia is fine. Maybe it’s only been abandoned because it does not have parameters. If that’s the case, an update would be awesome. If not though, that meant it’s been trained to its best form. Maybe they tried another update to it, but only got it looking the same or worse?
CUGAN is an AI image-processing framework developed specifically for animation. It’s based on cascaded u-nets, which are a form of deep-learning that was originally developed for biomedical imagery and then adapted to more important things like upscaling cartoons. CUGAN stands for “cascaded u-nets generative adversarial network”.
There are different CU models, but CUGAN is the most advanced. Sadly, most animation models are more advanced than GCG. Here’s a comparison of GCG and three CU models:
They are better trained for noise reduction and general enhancement. Gaia does neither.
4 Likes
Neither Gaia models do anything with noise—except maybe bake it into solid colors. The CG stands for computer generated, so that probably means it was trained on modern digital CG animations with no noise. It can do a really good job on those.
Looking back at the example use of the Gaia models in VEAI 2.6, Gaia HQ says it’s for high quality sources with an example picture of a real building, and Gaia CG says it’s for computer generated sources. I remember someone on these forums—I think from Topaz—saying Gaia was created for cartoons, but I don’t think that’s true. I’ve had better success with Artemis MQ on cartoons.
If I’m right about all that, then what you want is a dedicated model for older cartoons (or any cartoon with grain), and that is something that Topaz has not supplied yet.
You may already know but you can use chaiNNer along with CUGAN to upscale your files…
If you have an AMD or Intel GPU you can use the NCNN versions.
chaiNNer is my favorite application right now. It’s an incredibly well-designed app, and it’s allowing me to achieve many of the things that I have been working for ages to achieve with AI video processing – not just with animation, but with recordings and live-action as well.
ESRGAN has to be the world’s most advanced video AI processing framework. I didn’t realize that it performed detail generation like Gaia HQ, Iris, and Rhea, or facial enhancement like Proteus 4, Iris, and Rhea. Several of the Topaz models seem to be based on it, or at least had it integrated into their development – and, yet, it’s still more accurate than they are. Here’s an example of detail enhancement:
This is a scene from an animated production with a very small and not well-defined detail. Gaia CG, as usual, is completely useless. With Proteus, the object in question starts to look like anything at all. With Rhea, you can start to see that there are different segments to this item. With Iris, you get a better picture of what the item is – a circlet with a gemstone in the middle. With the animation variant of RealESRGAN, you can see the texture and shape of this circlet. And with the standard RealESRGAN, you can clearly see that the object is a metal band with a green stone in the middle.
Here is the entire scene for reference.
Kudos to ESRGAN. I do also make extensive use of CUGAN; it actually enhances the files unlike GCG, doesn’t produce artifacts like Artemis, creates a sharper image than Artemis or Nyx, and conveys fine details in animation more accurately than Proteus or Rhea. Iris also does well with animation, but CUGAN does slightly better. Sometimes it’s not as sharp as Iris, but I can compensate by using a sharpening filter or doing a double enhancement with CUGAN + a sharpening model.
4 Likes
Good work on showing the results of the models. I believe that definitely shows how necessary it is to update Gaia CG. I used to experiment with this model myself, and I was not able to get decent results either. In fact, I have not been able to recall the last time this model was updated at all.
I think you should rename this idea topic to something along the lines of: Make an enhancement model for old cartoons.
Gaia was not made for old cartoons. Any mention of Gaia being for cartoons must be in the context of a “cartoon” being a “computer generated 3D animation”. For that use case, it does what it’s supposed to.
1 Like
i like gaia for that exact reason.
it’s a reliable detailer and i don’t have to be worried about the erasure of fine patterns or noise.
Gaia HQ enhances detail (on a small scale) – and noise; I frequently found myself needing to use a second model to undo the damage that GHQ does and eventually stopped using it altogether. It is more destructive than beneficial, and is obsolete now that modern TVAI models (Rhea, Iris, Prob4) also generate detail, better than GHQ does. You can simply turn down or disable the denoising, dehaloing, and compression-reversion if you don’t want them to do it.
Gaia CG does nothing other than apply a thin film of antialiasing to the video.
1 Like
If I knew how to update Gaia to where it would be better able to keep up with demands, I would write about an update to the model since it’s clear that it’s been left behind.
