is there any Change Log or any details what was changed, enhanced in Proteus v4 vs. v3?
the latest release post doesn’t mention anything on it, only a line about Apollo.
1 Like
If a AI model isn’t in the GUI, then the developers don’t intend you to use it. Since Proteus V4 isn’t in the GUI, they don’t intend you to use it.
As for what Proteus V4 offers, the details have been provided for Beta testers and here are some notes I can share:
- At the moment Proteus V4 is still under heavy development and is still in the “Alpha” stage of development.
- Proteus V4 is given this “Alpha” tag because it hasn’t been optimised for performance, is still under development (as pointed out before), you can only do a 4x upscale (presumably 2x is planned for the future), and some of the settings to adjust Proteus V4 aren’t doing much or aren’t working properly.
- Along with all that, Proteus V4 has a tendency to reduce the detail in videos when compared to Proteus V3.
- Also, Proteus V4 can introduce various artefacts like blotches, halos, etc
Is there anything good about Proteus V4?
Yes, however I’m not sure I can share that with you since you aren’t a beta tester. So all I will say is that Proteus V4 is the model they reference in these places:
Source: Video Roadmap Update (Jan 2023)
Source: Video Roadmap Update (Feb 2023)
Source: Video Roadmap Update (March 2023)
1 Like
So why do I have it in my Models Directory (C:\ProgramData\Topaz Labs LLC\Topaz Video AI\models\prob-4.json)?
Many of the .json files from the Beta get left in the models folder in official releases. But the developers do not expose them in the GUI so most users don’t use them. Obviously, you can use command line arguments to activate them, but a vast majority of people won’t do that.
Why are many of them left there? I don’t know.
Examples of models that are in the models folder, but have yet to show up in the GUI of an official release include:
- Proteus v4 (prob-4)
- Chronos v3 (chr-3) - (Showed up in 1 or 2 betas, then didn’t get released. In the official releases we still use Chronos v2)
- Apollo Soft (apa-1) - (It was in betas, then development appears to have stopped in favor of Apollo v8)
- Apollo Sharp (apb-1) - (It was in betas, then development appears to have stopped in favor of Apollo v8)
- Apollo v7 (apo-7) - (I don’t ever recall seeing this in a beta, I think we jumped straight from Apollo v6 to Apollo v8 and yet there’s a Apollo v7 model file. Maybe it was from internal testing.)
- Apollo Fast (apf-1) - (It was in betas, and is in the benchmark section of the latest release, but you still can’t pick it from the GUI)
5 Likes
thanks for the detailed response. I wonder if I should try them via CLI… heemm…
You can try them if you want, most if not all of these should work.
However, I’m not sure if it’s worth trying Apollo Sharp, Apollo Soft, or Apollo v7 since they were discarded in exchange for Apollo v8. And based on my limited testing, Chronos v3 isn’t noticeably better than Chronos v2. But I may of not been testing the areas where it was improved.
Since Apollo Fast is in the GUI benchmarks, and is probably going to be released soon, I’ll tell you the general difference compared to Apollo v8.
Apollo v8 generates 8 frames, and then it discards the frames it doesn’t need to create the result you want (E.G. If you select 2x slow-mo, it will generate 8 frames, then discard most of them).
Apollo Fast is basically Apollo v8, but instead of generating 8 frames and discarding the unnecessary ones, it generates 4 frames and discards the unnecessary ones. So if you need to generate 4 or fewer frames, then Apollo Fast will give you similar quality as Apollo v8, but be roughly twice as fast.
There was a bug report in the Beta testing section that Apollo Fast produces weird star artefacts on some videos. I’m not sure if that was fixed, or what causes it. Or even if it’s limited to Apollo Fast. So be careful with it.
If you want to test Proteus v4 and compare it to Proteus v3, then I recommend that you compare Proteus v4 with a 4x scale to Proteus v3 with a 4x scale.
This is because Proteus v4 only has a 4x upscale option, and Proteus v3 does some stuff differently when it does a 4x upscale compared to a 2x upscale. So it’s best to compare the two AIs on “even playing grounds” by using 4x on both.
Note 1: Proteus v4 auto mode does work. It uses the Proteus v3 AI for picking auto settings to pick auto settings for Proteus v4. Since it’s not tuned for Proteus v4, the auto settings picked for Proteus v4 probably aren’t that great.
Note 2: Proteus v4 is still being trained, for the last three betas we got a new “checkpoint” (version) of Proteus v4. So keep in mind that you will be using an old version of Proteus v4 and won’t be seeing the latest benefits that we see in the beta program. Also just keep in mind that it’s undergoing rapid development and what you see now is almost certainly not indicative of the final product.
Note 3: Also, I’m not sure if the developers will take your feedback on Proteus v4 if you post it here in the “general forum”. Both because they may not be looking for Proteus v4 feedback in this place, and thus may not see it, and because you will be using an outdated version compared to beta testers.
2 Likes
I was planning to try Proteus v4, but I am only upscaling x2, not x4. the question is what is currently proteus v4 design on x2 upscaling? or what is the current expected outcome on v4 for x2 upscaling?
If you setup Proteus v4 to do a 2x upscale, then it will do this:
- Upscale 4x using Proteus v4
- Downscale back to the desired 2x using lanczos
- Save the video
If you did the same with Proteus v3, it will do this:
- Upscale to 2x using Proteus v3
- Save the video
1 Like
Strange how come, Proteus v4 can’t do x2 direct Upscale.
what was the reason for that that v4 can’t do x2 direct upscale, do you know?
2x upscale and 4x upscale are handled by different models. At the moment, Topaz Labs is only making the 4x variant of Proteus v4 available for testing.
Why would they only make the 4x variant available? My guesses would be:
- Since they’re rapidly iterating on the AI, they would prefer to work on 1 variant. And they picked 4x. They might of picked 4x instead of 2x because:
- It’s more versatile, as in you can upscale to 4x, then downscale to many other resolutions. This allows beta testers to integrate it into their workflow for testing easier (A lot of people are working with 720p or lower footage, and are upscaling it to 1080p or 2160p (4k), so having a upscaler than can 4x is important)
- With a 4x upscale, more pixels are being generated by an AI than the 2x upscale. This makes the unique properties of Proteus v4 more apparent, which allows the developers and beta testers to better identify what the AI is doing wrong or right.
In the future, I assume they plan to release a Proteus v4 2x upscale. But at the moment 4x is the only one avaliable.
3 Likes
got yah… (Thanks for the very detailed response BTW)…
I tested the Prob 4 (tweaked the Command and used CLI) and for sure generates less artifacts (or more precisely, dosn’t retain the artifacts that the original video had), at that area it excels much over Prob 3. but i noticed it is much blurry and less sharp then Prob 3, even when i bumped the Deblur and Sharpening to 100% it was less sharp & more blurry then my Prob 3 (in Prob 3 I usually work with Sharp=30 and Deblur=30). in both cases (Prob 3 & Prob 4) I use relative Auto and play around only with the Sharpening/Deblur.
Have you noticed the same or it’s just me. that it is more blurry then Prob 3?
P.S. I tested both Prob 3 & Prob 4 at the same output resolution of x4, with the same source file(s) of course (2 source files).
I used the Video Compare tool to compare them both (Prob 3 vs. Prob 4)
Video below is not related to my testing, it just to illustrate the Video-Compare tool capabilities conducted for my testing/comparing.
1 Like
Based on my own testing, Proteus v4 is much better at removing compression artefacts from videos, what I assume you meant about by “dosn’t retain the artifacts that the original video had”. However as you pointed out it is typically more blurry than Proteus v3. This seems to be because it has a really aggressive denoiser component that removes a lot of real details.
Although, based on my testing Proteus v4 can introduce it’s own artefacts. For example turning fine details (like fluff) into blobs, or turning out of focus backgrounds into blobs.
As for your concerns about the Sharp and Deblur sliders not doing much, that has also been an observation made by Beta testers. Many of the sliders don’t do much or do not work.
Once again, Proteus v4 is still in the early stages of development and will probably see drastic improvements before it releases.
It seem that you have joined the forum for almost a year and with good activities according to forum stats. If you like testing new features, I recommend you join the Beta Tester Group.
You might message Topaz staff @ida.topazlabs to let you join the beta group.
(I have also PM you the link to VEAI beta group)
Note: For joining the beta group, you have to include your computer hardware configuration into the Forum Profile About Me section in the preferences.
1 Like
It seems someone on the Beta program shared details on Proteus v4.
The main advantage Proteus v4 has over v3 is how it upscales faces. See:
perfect. I just wanted to know rather I was the only one experiencing that or it is a know current limitation. now it is to bee seen rather Prot 4 since it removes so much details, if it’s worth using over Artemis. Artemis is know to remove tons of details, Gaia on the other hand hardly does as it is meant for High quality videos that don’t need treatment as much. Proteus was some where in between.
now it might be that Prot 4 and Artemis would not be much different to each other at the outcome results. maybe worth wile for one of your Beta guys to run a compare test between Artemis and Prot 4. it could be that Prot 4 would just come out as just another Artemis like filter rather then Proteus family filter.
Any way thank you for all your help!!!
Yeah, that is what they claimed all along in the Road map, that they are focusing on face restoration.
On the clips I tested, it did not only restore the face but also other distortions I had on the human body that Prot 3 couldn’t do. so it did a fairly good job.
I will sure consider this… thank you for your gesture…
