Estimate and Second Pass

  1. Why does Estimate return the same parameters regardless of which model is being used? For example, why would the denoise values under Proteus be the same as Iris 1 when Iris 1 already has a higher level of denoising built-in?

  2. Does Recover Original Detail under Second Enhancement operate based on the source file or on the first enhancement? It appears to be referencing the first enhancement.

  3. Also, does anyone have or still use the Gaia 1-4 models? I’d like to see how they compare to V5.

1 Like

Iris and Proteus appear to apply the same estimates to produce different results. But if you switch enhancement to Nyx the estimate changes. So it’s not the same for every model.

So why do Proteus, Iris 2, and Iris 1 all apply the same parameters? And why does Theia not have an Estimate or Auto function? Theia retains the parameters that are set using Estimate/Relative to Auto on different model, but there is no option to do this on Theia itself.

Iris and Proteus may be variations on the same or two closely related bases. From beta build to beta build they tend to take turns being the one testers say is “better,” but the differences seem awfully subtle to me.

As for why some models don’t have the same adjustments, couldn’t say without knowing the inner workings of the AI, but generically speaking, presumably because either the developers haven’t figured out how to make them work or they have and the results weren’t good.

Theia was made before before Proteus and Proteus was the first one to have the estimate function.
At this point in time, it feels like, Theia is only available in case it worked well for you in the past and you need to get those same results again.

It would be nice if the developers provided some explanation of how the models and other features are built and how they function.

For example, the values used for Auto seem to just be more maximized versions of the values returned by Estimate, whereas Estimate appears to return values that are more intermediate. And the results produced by Iris LQ, MQ, and Proteus are all very different, so I think the engineers should look into adjusting how Estimate functions on these models.

Some explanation of Theia’s intended use-cases would be nice, along with some explanation of how Theia and Gaia are built. Gaia functions more like a filter than a genuine AI model – there’s a filter I use in a different program that does exactly what Gaia does, with the added bonus that you can control the intensity. I’d to see if the older versions of Gaia behave any differently.

@ForSerious @eh117 Is Thaia mission statement (primary objective) is to enhance a video at the same resolution or for upscaling with enhancement? what is the original intent of this AI?

“Upscale or sharpen input video with more detail, with parameters for fine-tuning.” and “Upscale or sharpen input video at higher fidelity, with parameters for fine-tuning.” Those are the tool-tips for Theia in VEAI 2.6.4.

@eh117, For over a year, I was also under the impression that Gaia was basically the same as a Lanczos upscale filter. Not long ago, I did some critical comparisons and Gaia is able to create more accurate details than just a filter. I think it’s better than all the other AI models at leaving everything it doesn’t know about alone. It’s weakness is that it tends to ‘bake’ film grain into things. So like, a pure colored wall will have little dark spots on it after Gaia.

1 Like