You can try IDM.
The download support multi connection.
In my computer, I can get up to 69MB/s and total download time is around 2~3 mins.
thatâs right quality more important than speed.
I rather have a model that take a minute to process one image but give a feeling of enhance command
(just like in the movies)
1 Like
You have played around in the folder of the model files.
This is probably the problem.
I would reinstall it.
And delete the whole tgrc folder before install.
@TPX,
i deleted the whole installation including tgrc folders.
Did a complete fresh install.
-
Tried processing with GAI optimized model by letting GAI download the models - Same blurriness.
-
Tried copying CPU only model as suggested by taylor - Same result.
The results are worse than 5.4.5 in all cases.
Someone let me know if i m doing something wrong or if you see better results than 5.4.5. otherwise there is no need for an upgrade.
Also the fact that they have removed the OV options in preferences says a lot of about where GAI is headed.
5.4.5 will probably will be my last buy. I donât expect it to get any better than 5.4.5 and i m saying this after spending hours and extensive testing.
1 Like
Could you show a comparison?
Sure, but this time i suggest you give me a test file.
I re-installed 5.3.2, one of the last builds where Iâm sure that I can use CPU only, as we can manage this in preferences.
Then I tried with last beta before 5.5. I think that this would interest @taylor.bishop too.
- Yes the results are different, and yes I disabled âmodel downloadingâ so I just have models packed with the installer.
- @zamfir1235 : I tried with a single picture, a cartoon in CGI and I can say this : by comparison with the same picture done in CPU only, sometimes old model is better, but most of the time new model is better, frankly. Letâs say that 80% areas I checked are better now. Checked with a new installation of 5.5.
- The only thing I didnât know yet is to check that *-tf.tz models produce blurry results. Iâm downloading all models, thanks for the share, but 17 Go will take me at least a day, I donât know if Iâll succeed.
- I also checked that 5.5 models are the same that the models in beta. Itâs OK.
I updated to v5.5.0 and thought the software doesnât work anymore, as it initially got stuck at âDownloading optimized modelsâ at 2%. However, it took 20-30 seconds and once everything was downloaded, it worked. I guess Topaz has to bump up server capacity to enable quicker downloads! At the moment, System: Ryzen 9500, 64 GB RAM, Win 10 Pro, 240 Mbit connection.
1 Like
Try to use the 192 Models only.
Will try that later this day or tomorrow.
My only request to Topaz devs it to make OV optional and NOT mandatory.
OV has been optional all this while, why make it mandatory?
1 Like
Can someone explain what is 192x192 and 384x384 models?
Which one would be better?
So i understand that the 192x192 or 384x384 is the block size.
I have now tested without ov. i.e only CPU mode in 5.5.0
Model file used ggn-v1-fp32-192x192-tf.tz. This uses only CPU and block size of 192x192 from the logs.
Observations
- Time taken to process is same as time taken in version 5.4.5 CPU mode
- Lots of parts have blurriness with ggn-v1-fp32-192x192 as compared to the 5.4.5.
- Changed block size to 384x384 , no difference noticed.
Model used : ggn-v1-fp32-384x384-4x-ov.tz
Observations
- Almost same time taken with OV on in previous version of 5.4.5.
- Image quality with less sharpness overall. i.e image quality is similar to OV ON in version 5.4.5.
Anyone can make these observations with given model file above.
So OV on makes image worse in quality but increases speed. This is not necessarily a bad thing as it depends on the requirements of the user. But the same speed can be achieved in previous versions. Nothing new in this version.
OV does not produce a sharp detailed image which is what is required in a up scaling photo software. Donât know why anyone would want to use it.
The new version 5.5.0 even with ONLY CPU model ggn-v1-fp32-192x192-tf or ggn-v1-fp32-384x384-tf produces less detailed and blurry image compared to the CPU only (giga.tz) model of 5.4.5.
In summary,
I have tried OV, TF models with different block sizes and the newer 5.5.0 produces less detail and less sharpness compared to the previous version 5.4.5 models.
The speed noticed by many is only due to OV models, as OV option is not present anymore to disable it. If you disable OV, there is NO speed improvement.So i fail to understand how topaz claims there is a MASSIVE speed improvement, when there is NONE.
All this is evident from the logs for anyone to see.
So is it now just soft and blurry like a simple bicubic upscale in Photoshop, or is there some Topaz magic still left to be found in the result? (Iâm waiting before I pay for the update, so I donât waste my time and money on a step backwards.)
I donât have any idea about photoshop upscaler.
GAI 5.4.5 is still better if you want to put your money in it.
But i would definitely not pay for 5.5.0.
I donât understand what you talk about.
We both were in the beta.
Me and also others in the beta said that the models are now better and much faster than before, we went down from minutes to seconds, from hours to minutes.
Also other people in this thread did report that this software is better overall now.
So whats up here.
What and how did you enlarge and which hardware are you using.
And please show some samples, this is about images not magic.
2 Likes
I mostly use GP AI to upscale low-res JPG scans of old colour slides or degraded JPGs from my first digital cameras of some twenty years ago.
If anyone sees a real difference between CPU and GPU rendering in v5.5.0 and the one on the right which I created in an earlier version of GP AI back last Nov, then they have better eyesight than me.
Plus the fact that 5.5.0 is quite a lot faster. Not that that worries me. I have all the time in the world!
Time and again i have shown proof. How much more proof do you need?
I donât think people have tested it enough. Beta testing was not sufficient.
The models are not better than previous versions. This is my observations, yours may be different. This is because image quality assessment is subjective. We could use objective methods to test e.g PSAR and other techniques. But that is not part of the testing process anyone follow here. So that would not be any better.
For me its not worth it and i do not see any improvement over previous version and i have been comparing newer version whenever they come out.
i do not need to show any samples as that would only induce bias and cherry picking.
I have reported my observation , you can do the testing yourself and see for yourself.
if you are happy with the result , then it does not matter what i or anyone says otherwise.
Theyâre all supposed to produce the same output. There is a block size that for your particular system, is âoptimalâ in that it will result in the fastest processing. In other words, itâs simply a performance thing. Not that big a deal if you process only a few images, but if you have hundreds or thousands you want to run through via batch, it will save some time. Obviously this is a bigger deal in Video Enhance AI since a 90 minute movie for example will have around ~135,000 frames (or individual images).
Before processing, the application does a test (it doesnât need the model files for the test) to determine which block size results in the fastest processing, then, it downloads that block size for the AI model youâre using (ox, ov, etc.).
1 Like
Then there is nothing to talk about.
Itâs like when you buy a car and the seller doesnât want to show you the color because he thinks you could make it bad. 
so you are saying we should not report any observations or opinions?
Besides , i have shown you many samples before. but you never acknowledge them , nor do you care to report your observations about it.
I think you are happy with the results and that is fine.
