Topaz Video AI 6.0.2 + 6.0.3

Yes please! We’d love to get a comparison thread going between our current models, other models, and our upcoming model which is going into beta this week.

3 Likes

First, it’s nice to hear that there’s a new model, so finally added functionality and not just messing with GUI and such.

But, I’d prefer refinement of the existing models (especially Iris V1 and 2 LQ and Rhea) instead of new, but also likely somewhat unpolished models.

E.g. We last had an update to an existing model with Nyx V3 I think but it’d be extremely nice if the generally good working (but with known flaws) models Iris and the promising but also quite unpolished Rhea got improved/refined.
Artemis is even at V12…

We’ve thought about this quite a bit, and this is my current perspective on model versions and naming:

There is only a finite amount of potential for model iteration, and I think a lot of our users think that we are making a choice between exploring a new model that attempts to perform beyond the limitations and weak points of a previous model, or trying to smooth out specific artifacting issues or limitations of an existing model.

For example, if Rhea was released as Iris v3, it probably would have been received differently but would still be the same model. Same could be said for Nyx, which easily could have been released as Artemis HQ v13. However, we think the progress and technical differences between Artemis and Nyx

In other words, the naming and labeling of models is something we’d like to simplify, and with the upcoming model we think we have made progress in simplifying the choice of selecting a model and knowing what it could be specifically good at processing.

6 Likes

Good time of day. I ran into this problem: after processing the video, it goes in jerks. This happens mostly with long-running videos (concerts, movies, etc.). What could be the cause and how to solve the problem? Has anyone else encountered this?

This is exactly what I thought when I first saw Rhea - it’s kind of an updated Iris with a bit of Nyx in it - so yes, I can see your point.

2 Likes

Fortunately it wasn’t and give us the ability to pass Rhea (standard) + Iris to get something “good”

That’s a sensible reasoning, although the problem here becomes another, for me:

by “continually” releasing new upscaling models with “new” names, it starts to get a bit confusing for those who have to “choose” their own appropriate upsaling model for their source video, and so (in my opinion) one should “start from scratch” and choose “basic” models chosen according to the type of source source, and eliminate “duplicate” ones: for example, from what I have seen, the basic “Rhea” model (the Rhea XL for me is unusable to date) is a kind of IMPROVEMENT and evolution of the ‘Proteus’ model… but at this point I wonder: what is the point of keeping the “Proteus” model?

Obviously they remain tendentially “different” models, but I hope the concept is clear!

1 Like

We had some reports of dropped frames or framerate errors recently. The latest release 6.0.3 should resolve most of these cases. It’s also possible that we will need to remove duplicate frames from your input video before processing.

I recommend emailing help@topazlabs.com for direct help with your files.

That’s exactly it – I think some of the requests we get for “Nyx v4” or “Artemis v13” or other version number updates to existing models are really just different ways of saying “Try to get better results for {X} type of video”.

Our upcoming model is a major step towards simplifying the model “toolbox” and having a more intelligent process with less fine tuning required.

3 Likes

I think I know what you’re trying to say, but Rhea has always been defined by the special face enhancement, were Proteus doesn’t have that. I think the difference becomes less clear when comparing Iris to Rhea. But then, Iris is supposed to work on low quality sources, where Rhea is not. It would be so much harder to remember that kind of difference if they had named it Iris 3.

1 Like

But when Rhea is simillar to Iris, why Rhea is only 4x optimizeable and Iris not? When NYX is in Rhea included (what I’ve suspected for a longer time) NYX is also not only 4x optimized.

I could imagine it would be difficult to constantly update all the models, even if the basis may be similar. Would the long-term goal be to merge models so that there are fewer but better ones a desired step?

I work very often with Iris and Rhea while people say RheaXL is unusable :laughing: Maybe I’ll make a comparison video about it. I see potential in the combination of RheaXL with Iris, which often provides the best results when source is not very good, by overlaying RheaXL encode with Iris encode (video editing, set opacity) and then some finsihing steps into TVAI.

This overlay weakens the partially exaggerated face reconstruction of Iris and partially compensates for the lack of detail in Iris and covers up the “roughness” of Reha (what Rhea included NYX does)

Maybe something like that could be a way to got with TVAI, internall as a second step an adapted iris model does a second encode and “fusion” this with Rhea encode

2 Likes

I mostly use Iris medium quality to be honest. Reha is too slow and not worth the extra time and power consumption by doing worse than Iris medium in comparison often. Reha XL brought the worst results to me while taking a serious amount of time and power also. I won’t tell about the annoyance of the 999th GUI change anymore as obviously nobody at Topaz cares about this anymore anyway.

3 Likes

Yes I tryed medium Iris, but for me it seems it’s not on the level as Iris LQ V2, but Iris LQ is not really made for good sources. You can trick it a little bit by presharpen source (but then also you may loose some fine details) and max out deblur levler of Iris LQ, but that is suboptimal, we need a Iris MQ V2

I know RheaXL is very problematic, but for some scenes, when you overlap it with other model then low to medium artefacts RheaXL produces are not visible anymore…and then you can get results no other model can do, because the ammount of keeping detail RheaXL does is unique

6.0.3 just crashed i dont know why.
Device: M4 Max 128Gb

1 Like

We’ll need your logs sent to help@topazlabs.com for a closer look:

To gather logs, please select Help > Logging and make sure that there is a check mark next to File Logging. Next, recreate the issue and then return to the Help menu and select > Get Logs for Support and attach the “logsForSupport” file to your reply.

If you cannot recreate the issue, the logs folder probably still has the recent crash captured.

Does 6.0.3 include the models mentioned in this article about Project Starlight? If not, is there a beta I can signup for?

2 Likes

Starlight research preview is available right now in the Video AI 6.1 Beta. In the next day or so, we’ll be adding Starlight to the Topaz Labs web app for easier access.

Here’s a direct link to download the 6.1 beta if you’re not already on the Beta forum page:

6.1.0.0.b

Thanks Tony! We only using the main app rather than the web app. I’ll download & start testing the beta.
I assume the beta can be installed concurrent to the existing Video AI app, correct?

Yes, the beta version is installed alongside the standard release. You’ll have two apps “Video AI” and “Video AI Beta” that do not share any files