Discussion | Wonder Discussion and Feedback

[zelenooki87 (Zeleni)]

Let’s be honest here — Wonder v2 isn’t quite performing as expected, even on smaller images. From my testing, the 1x mode is the most usable since artifacts are nearly invisible, but at 4x the quality drops noticeably. Wonder v1 actually produces better results at 4x, though it has its own artifact issues as well.

What I think we really need most is a properly working Recover v3, and — perhaps even more importantly for those of us with capable hardware who are currently experiencing issues — Wonder v2 delivering pixel-perfect results that match the quality we see in the web render.

The honeycomb artifacts with Recover v3 still haven’t been addressed in either Photo AI or Gigapixel. These have been present since the Gigapixel beta. It’s also worth noting that Photo AI 1.3 didn’t go through a beta testing phase at all. Given the significant under-the-hood changes (the neuro server architecture, etc.), I personally feel a proper beta period with user testing would have been really beneficial. That might explain why we’re seeing such a large number of reports about model loading issues and quality problems with both Recover v3 and Wonder v2.

Another concern is that instead of releasing a patch to fix model functionality (such as the .dll fix that some of you have mentioned), the team hasn’t put out either a patched version or that specific file. There also hasn’t been much official acknowledgment of the artifact issues, which are quite noticeable. Many users, myself included, have submitted detailed examples.

I appreciate everyone sharing their findings here — hopefully this feedback helps move things in the right direction.

[TPX (Thomas K.)]

Did you try to make images smaller before upscale with Wonder 2.

From 6K i make images smaller to 3766px and upscale them with wonder 2 (cloud 4x) and i get a very good result.

[yabino974]

I completely agree. To me there should be a model between Wonder 1 and Wonder 2, aimed at high res pictures, or phone pictures.

I process ALL of my phone pictures with both Wonder models and blend them in Photoshop. I do it with x1 scale. It smoothes everything out nicely and does a great job at adding definition.

On faces it’s sometimes too much, so it would be great to have a minor/medium/major enhancement mode with a preview like with Super focus.

On a side note the “Recover faces“ would benefit from generative AI to add back lost details and add more sharpness instead of smoothing/blurrying things out.

[Ange.topazlabs (ange.topazlabs@gmail.com)]

@zelenooki87 - For the artifacts: We are working with 2 users that flagged this to get to the bottom of this. This is not something we can replicate with any machines we have and we have screenshare booked and communications with them to see and test steps with users that have this. Things we cannot reproduce will take time to troubleshoot and fix and between Photo and Gigapixel, there is a total of 5 users that have flagged this. Once root of issue found, hopefully we can have a workaround, and if its an open item to do with the development team, issues are then tackled in order of users that have flagged the issue.

The DLL fix, had to be tested with multiple users as some users were still flagging issues. Once fully tested and confirmed working it is worked into a patch that is set to release this week. However, all users that have flagged it already have had the fix sent to them, and it allows them to be on the last version. A patch would not change anything as they already have the fix. The patch is aimed to be released tomorrow, but all users that flagged the issue already have fix and v1.3.1 working.

Wonder local vs Cloud: We also have the development team work on better optimizing the Wonder local model to match Cloud, but it is a significant smaller model, and it will take time to optimize such a big model. You can use the cloud as its unlimited and free.

Also, while your files were relayed to the development team to help them see unwanted results, the “original” sent has a lot of RAW noise and hot pixels in the file, flattened, and comes from Photoshop, and it would not be the real original. It already has flattened issues into it. You can DM us the real original to help us suggest the correct workflow for this file, which would be to use RAW Denoise, not Wonder. The Wonder model is to fix compression artifacts in low-resolution files, which wouldn’t be the case here. Right now it is fighting with RAW noise artifacts being flattened into the file in Photoshop and causing extra artifacts to remove. If you send the real out of camera file we can test it to suggest best models to use to get a clean output. A full resolution file shot with your Canon Powershot camera would not be a case for Wonder. It could also be a non-RAW file, but with lots of noise. If shot JPEG - We would still need to remove this noise with Denoise with the original file, and use proper models for the file, which wouldn’t be Wonder as it’s a full size. full resolution file shot on a Canon camera.

We do have other users that shared real cases for Wonder and shared inferior local results compared to Wonder Cloud, and that was also sent to the development team working on closing up that gap.

@yabino974 - we discussed by DM, and while you can use the model on files not meant for the model, yes it would require blending in PS and give inconsistent results as using both models outside their intended use. Glad it can help your workflow though.

——————–

I see all 3 of you added your vote on for the feature request for a Wonder-like model for higher resolution files, that is great. It would not be called Wonder, however, as it would be a fully new model, trained for a different purpose and different image set than Wonder is trained for.

[JEM]

I’m testing the new Wonder 2 model, and for high quality portraits, it’s probably the best model out there right now, but for everything else, it’s actually worse than the Wonder 1 model. This is a crop from a photo taken with the phone, with some noise and processing applied by the phone itself.

This is Wonder 1:

This is Wonder 2:

[Fotomaker]

The top one looks like a screen effect & the bottom one, paint dabs.

This is so nostalgic to when Topaz filters/plugins offered creative effects.

[alexandre.topazlabs]

None of the results are good in my opinion. Do you have the original file for us to test? Maybe another model would work better.

[JEM]

[AND-E]

Hi Jem.

@JEM

This is a crop from a photo taken with the phone, with some noise and processing applied by the phone itself.

That’s your issue you’re using a Phone from a distance and you’re shooting though Atmosphere.

Your Phone, has a tiny Sensor which has already been highly processed before hand and this is the result of Applying Wonder to that image beside the whole Image probably still looks good but, Cropping it this tightly will always reveal imperfections.

All these AI Models, work better with Great images and less so with poor images, similar to Film restoration.

Restoring an old Film or TV show will always render better results than restoring Home Video it’s all about the quality of the original material.

Wonder is designed for low resolution images with compression artifacts but for best results it still need something good to work with and for your image both Wonder 1 or 2 has done a reasonable job with the source material.

Hope this helps

[jo.vo]

Not really the original anymore,though

[Ange.topazlabs (ange.topazlabs@gmail.com)]

You would need heavy creative generative models to recover this, like jo.vo posted. It is not a file Wonder can salvage. At Topaz such models would be Bloom or Topaz Gigapixel’s Redefine.

[drdre1411]

Although my usecase is not technically what these models are for ( scans of workbooks for a school) I do think a mix of Wonder 1 and 2 would work very well. 1 does a great job at removing smooth edges but does not seem to do too wel with text, and it stumbles on some noise, while 2 actually usually removes all the noise properly and does an excellent job with text, just has a lot of issues creating artifacts, and creating harsh jagged edges.

And actually Dust&Scratches does a great job too, but seemingly also stumbles and removes certain text and line objects.