Video Enhance AI v2.2.0

Topaz Video AI Releases
332

Hi Taylor, congrats on the effort you’ve put into these enhancements. This is moving along very nicely to a more workflow-based model that I suggested a while back. I’ve had a quick experiment with it and so far it’s looking great. Two observations for your consideration:

  1. Still as in all previous versions, the very first time I attempt to Preview render after installing the update, it instantly crashes out. From the second launch onwards, Preview is fine. (Currently Windows Pro 10 x64 18363.)
  2. After you’ve watched Preview, when you select ONE preview pane and change the model, clicking Preview re-renders ALL the models so you have to wait for three even though you’ve only changed one. Would be better to keep the other models in cache so they don’t have to be re-rendered, to speed up workflow.

This multiple side-by-side Preview makes it SO easy to select the best model for each clip. Great job!

333

OK, you know where your model folder is? Show me the “4K” model. Show me the “450%” model. Nobody else has them, so I would like to know what they look like.

Hopefully Taylor or someone will come along to disabuse you of the notion that there is any such thing as a 225, 450, or 4K model.

When I select 4K, my banner does read 4K. When I choose 450% custom, it reads 450%. No ERR, which is something only you have.

I am sorry that I never paid any attention to the ribbon location preference, my bad.

334

ERR
ERR2398×346 118 KB
models

As I can see, there’s a model “none” there that came from nowhere (I had never checked the models folder until now). But about ERR, I don’t know anything. All I know that it “gives” me lots of advantages. When I choose 4K, it is 4K. When I choose 4K UHD then 450%, it shows 450% like everyone one else.

Like I said. Anyone who can test just try and let me know. If you don’t have that option, then don’t mind it. My main post is just a question (with a question mark right on the title). It’s not a claim or something. But I have shown the proof. Maybe only people who have Ampere GPUs will have that option? Or maybe it was an error on my machine? But since this error is too GOOD, I doubt that it is a new hidden feature that I accidentally unlocked? That’s why I asked the developers to give me an explanation for this. No need to be grouchy about it.

335

Those are the json files, not the model files. Since they are text, you can open and see what models they use. The none.json (which has always been there in every version) shows that it is “Fast Resize Development only” and it has no models for fp16 or fp32.

If you look in the logs when you see that ERR, you will be able to tell which model was actually used, and I will bet you will find that it is 1x, 2x, or 4x, as there is nothing else.

336

When I use 4K UHD 200% with 4K output, it uses this model (ahq-v11-fnet-fp16-576x672-4x-ox.tz)

When I use Custom setting standalone with 200% output, it uses this model (ahq-v11-fnet-fp16-576x672-2x-ox.tz)

And ERR log file is just empty. The information above I got is from Main log file.

337

Interesting. You do have an oddity - which is it using the 4x model for a 2x resize in the one case. Certainly looks like it is using the wrong model.

The ERR log would only exist if there was an actual error, in which case you would not have completed the upscaling. What I was interested in knowing was what model was used when you actually saw ERR on the model overlay, which from what you said was at 450%, not the 200 used in your examples just now.

338

I have ERR on every scaling model at any % custom setting unless I choose SD HD UHD presets first and then Custom setting.

340

As far as I know (I’m not the dev), h265 won’t be available anytime soon due to the license. AV1 is new and it’s extreme slower than h265, only Youtube should use it because they have super computers. ProRes 4:4:4 is pretty heavy, you can try TIFF 16bit, and then combined the image sequence into 4:4:4 ProRes output video. Anyway, you requested some kinda heavy formats that I’m pretty sure you will need to spend days to upscale. TIFF 16bit is the best choice for you now, so you need to invest in Adobe or Davinci Resolve to combine image sequence into the output your choice.

341

Doesn’t ERR means the program switch to CPU instead of GPU for some reason ? That could explain the output differences. Is the speed the same ?
But like all others I’m not able to reproduce it from what you are saying and I have an Ampere card.

342

Still GPU with same speed anyway. It’s weird because VEAI uses wrong models on my end. Instead of x2, it uses x4, therefore the quality is really bad unless I did the method. I think some people may have the same issue.

343

Me too, I cannot “force” my system to show Err like Victorz3008 describes.

344

Thank you. I shall compare with the method I use together with NLE to downsize the footage to 1080p and write my opinion.
My speed is 0,90s/frame, I am very suprised the difference is not as high as I expected. I have only 6GB GTX1060 GPU and 16GB RAM. Very strange, I believed your system would be ten times or more quicker…you mentioned RTX 3090 in SLI…
PS Could you also check the speed of application of Artemis HQ 100% denoise/deblock on the sample downsampled to 1080p by any NLE you use starting frommy ProRes?
So firstly from my ProRes by 450% using AAA9 and then to 1080p by any NLE you use to ProREs or DNxHR and THEN cleaning this 1080p footage by AHQ11 100% denoise/deblock to ProRes in Topaz, I wonder about this speed on your system.
Thank you in advance.

345

Does it matter if you use TIFF, mp4 or ProRes output? I mean ERR abbreviation displayed.

346

It happens with any output format. I always clean install VEAI versions. No modifications or whatsover.

347

Usually, I don’t use 2 models on the same footage because it makes no difference at all. Also the speed is very depends on the first model you use. If your PC takes 0.9 and mine 0.2 then it’s about 4-5 times faster already. So I only need about ¼ time of yours on any footages. That’s a lot already.

348

I was thinking about that, and there was one piece of information missing in trying to understand your example from the logs: What were the dimensions of that original video you were upscaling UHD/200%?

349

Full HD 1080p.
I personally use 720x480 interlaced 4:3 DVD and 1920x1080 progressive 16:9 Blu-Ray only. So these are like the standard formats anyway.

350

I was finally able to get some of the results you see, and I did that by turning off “Lock output size” which otherwise takes a 4K setting/200% and does some really strange stuff. How is it set for you? For 720x480 set to 4K/custom 200% and output size unlocked I wind up with a 1440x960 image in the middle of a black 3840x2160 frame. When it is locked, 4K/200% custom is the same as 200% custom without selecting 4K first.

Incidentally, 720x480 is not 4:3, it is 3:2. I assume non square pixels, changed by the display/playback?

351

OK I was able to display that ERR after switching from 4K to custom without doing anything else.
So I did a quick test with some PNG output - 18 frames - with the 4K and ERR using Artemis HQ + LQ v12.
But sorry it’s the exact same output bit per bit. I run a duplicate file detector program and it returns all 18 PNGs as duplicate files from 4K/ERR processing.

352

I really have no idea because this is definitely a bug. I can’t explain why. It maybe better, or the same I don’t really know. But based on my testing, it’s much better because VEAI uses the correct model as it should.