We are working on improvecd ETAs and progress for these new models. The first run is generally inaccurate due to model downloads required, unzipping, or loading the models. Hopefully the runs after that are higher accuracy.
The large underestimates persist after several runs of the same model.
2 posts were split to a new topic: Wonder 3 Pink preview and ouotput on Windows
Yes it is.
Did take 25 Minutes to render each step.
There was an Denoise Max step too but i did not show the result.
A RTX 5090 would be, in theory, 14 minutes faster (40%) but thats just guessing.
Brought the Ryzen 9950X3D2 and it just feels great.
I unfollowed Topazlabs on Instagram. After they posted a video of a monster ballerina with a batâs nose for a head, whoâs anorexic and dancing through a forest.
Thatâs where I draw the line.
I donât want to be associated with something like that.
I saw it, very creepy.
Superfocus 3 on RTX 3080 TI (12GB) hitting 99% vram during pre-processing. Appears to be loading model over and over again. Hits 99%, then 14%, then 50%, then 99%. Is this a bug or a feature?
Done, thanks Ange for pointing out the âideas forumâ, wasnât aware of that.
High Fidelity v3 exhibits excessive smoothing; any input image defaults to a strength of 75, and even when manually reduced to 1, the result remains smoother than in HF v2.
source
HF V3
HF V2
That would be really good to know when that is fixed.
Hi Jason, this bug will be fixed today. We need to test a little bit.
Hifi V3 looks flat and like AI.
I also noticed it destroys details/smooths everything over. Not a model I can use in that state.
can you send the image used and your app logs (Help, Open Logs Folder, send all files)
to support@topazlabs.com and we can see what is happening in the logs and attempt to recreate it with your file.
Hi jason, The black border bug has been fixed. Please give a try.
Tried at x1.5 and 2.5. Cloud processing. No longer black borders. Thank you so much for sorting this out.
Much appreciated.
This is Wonder 3 - 2X.
Iâm not happy with the result, but the model is not made for this big image.
Made with an EOS R1 - using eye control and Pre-Continuous Shooting.
2 posts were split to a new topic: Denoise Max âAlways use this selectionâ - No way to revert
Once the calculations are completeâespecially with large imagesâyou must save the results without moving the cursor first.
Otherwise, the results will be lost.
Loading the new Neuroserver models takes a very long time, but with Wonder 2, itâs really fast.




