It’s just taken me over two hours to download the models, at an average speed of about 1.3 MB/sec. The apps themselves (and anything else I download) usually come down at about 15 MB/sec. What could cause this - and could it explain why installing the models into the app takes so long for some of us?
I’ve just tested something :
- Upscaled a picture x4 with models *-v1-fp32-384x384-4x-ov.tz. Downloaded via the application.
- Downloaded all *-v1-fp32-384x384-4x-tf.tz shared by @Red above, and used same settings and same picture. → So CPU only, as you said.
At first I just wanted to check if there were any differences, but I didn’t that because the result with .tf.tz was pretty blurry. Did I do something wrong ? tf.tz seem to be pretty useless.
Hi Taylor,
I have only -ov.tz files and i deleted them. But now when i try to process any image, it crashes.
I do not have -tf.tz files.
I m downloading the 15GB model files. Does it have the -tf.tz files?
It does. The version next week should have these pre-packaged for the future.
Can you tell what the files/models mean in the zipped model files.
e.g if i want to use only CPU, should i only copy files with -tf.tz?
What about GPU?
It would be nice if you could explain what the nomenclature is.
The answer to that is chunks, or rather how many are used to download a certain file. Gigapixel in it’s current state only uses one chunk while fetching for models from the server hence the painfully slow dl speed. So it makes sense that the only viable solution would be to implement multi-chunk download/parallel downloading within the program (a trick used by various download accelerators for ages), which would consequently allow your network speed to be used at full capacity.
1 Like
@taylor.bishop : I comeback to the Beta. I checked with a picture done with the beta, and the result is way better with the beta. I don’t know why, but for moment I don’t use the released version anymore.
Edit : Small question : during the beta, I was using this ***-v1-fp32-192x192-4x-ov.tz
That is different to what I use with released version, which is 384-384. Does it mean smaller is the number, better is the result ?
Edit 2 : After other tests : the one which gives better results (at last for the setting I tested) is the 192-ov.tz of the release. I think that I copied the files from the beta, in order to avoid downloading them.
Ask me questions if I’m not understandable.
I don’t think that explains why downloading the zip file was so slow. It was downloading at about 1.3 MB/sec, then I paused it to download the latest beta which came down at about 15 MB/sec. When I restarted the zip file it continued at its snail’s pace.
1 Like
Try internet download manager, i m downloading @ 12 MB/sec. Expected time to completion is about 20 mins.
Here’s me downloading the zip file with IDM using 8 chunks:
Now, same thing but with only 1 chunk:
I have reasons to believe that “model.topazlabs.com” i.e. the new server where the gigapixel models and the aforementioned zip file are hosted is more prone to slowdowns using only 1 chunk per download, whereas “https://downloads.topazlabs.com” i.e. the main server isn’t in the same predicament since it’s perhaps more performant/more optimised than the former, but I could be wrong.
I tried IDM - seven minutes @ 28 MB/sec. Thanks guys, I really appreciate it.
1 Like
I m now using TF to process the images after downloading the using only the -tf.TZ model files.
I have a processed 5.4.5 image and its blurry in comparison.
This is not good.
I get 14 MB/s when download the models with Edge.
eh! it went down to 1.4 MB.
I m reverting back to 5.4.5.
None of the models are coming even close to what 5.4.5 was processing.
Oh well, it was a nice try.
You spoke too fast hahah.
1 Like
That’s what happened to me - it started at 15 MB/sec then dropped to 1.3 MB. And stayed there.
Did that too to Edge, thx for the tipp.
1 Like
It seems that I’m not the only one.
yes, i think if there a update or an upgrade, the newer version should produce a better image than the previous version.
It should not be a downgrade at the expense of speed.
I could get the same image with 5.4.5 with OV on with faster processing time.
I fail to see what the upgrade/update is in 5.5.0.
I don’t understand how this is acceptable.
1 Like
The whole point of Gigapixel AI was to produce the best image from source without regard to speed.
Now , the direction has changed to producing mediocre image at faster processing.
I m really hoping the direction will change…
6 Likes
