I would say:
Wonder V3 is too artificial and make plastic Faces….
Wonder V2 is looking much more natural!
I would say:
Wonder V3 is too artificial and make plastic Faces….
Wonder V2 is looking much more natural!
Thanks, I responded to your email. We can continue the conversation there!
Please send logs from the Help > Open log folder menu and I can check what is happening with your authentication.
I reproduced this and gathered some logs to figure out the cause with the research team. Thank you for sharing this.
Same problem here. Its also using tremendous amounts of ram… running wonder 2, currently free: 2.7 out of 64gb. Vram only 20% utilized. Same thing happened using Wonder 3 which completed but results did not look very good. Using wonder 3 does not appear to be using gpu, but cpu. Vram stayed at 23% (on 12 gb RTX 3080 card)
Wonder v1 works fine and uses gpu, looks better than wonder 3
2026-04-26-11-52-46.tzlog (226.3 KB)
2026-04-26-12-13-5.tzlog (32.4 KB)
2026-04-26-12-26-20.tzlog (29 KB)
2026-04-26-12-15-14.tzlog (41.6 KB)
Thanks, I was just about to report the same thing. There’s no difference between the three presets, and I’ve got some noise creating patterns that remind me of the early versions of Gigapixel.
My GPU is a RTX4080.
Edit: Tested with CPU processing and same problem.
Denoise MAX takes forever to do anything after you press start.
After that, it finishes pretty quickly; something needs to be optimized here because the system seems to be idling.
The output from Wonder 3 was so bad that I didn’t save it (medium).
It looked like it had been processed with a 3x low-pass filter.
Wonder 2 throws an error. ![]()
Yup same here. I “wonder” how v3 is an improvement in any way over v2…
Hifi V3 swallows microcontrasts.
Hifi V2 is better for processing afterwards.
Hifi V3 looks too artifical, reminds me of AI and not a photograph.
The most useful result with my testimage was Denoise Max & Sharpen + Denoise Standart (background).
I could imagine to combine Denoise Max & Wonder 2.
From 1.4 can’t remember if it was 1.4.0 or 1.4.1 but now the wonder-2 folder has vanished.
I have mixed feelings. I’m getting really good results in some photos with the new Wonder 3 model, but in others I’m seeing noise patterns similar to those in the early Gigapixel models, which makes the images look very poor quality. I think there’s something wrong, but if it’s fixed, it could be a perfect model
Finally got Wonder 3 to work. Upgraded from 1.4.1 and couldn’t get the Wonder 3 model to run. Just got stuck on processing, even left it overnight. Uninstalled, deleted the Topaz Photo folder from ProgramData, Local and Roaming App Data. Installed 1.5, still had the same issue. Used Revo Uninstaller which also removed the registry entries alongside all the related files and folders. Tried again, and it seems to be working great. A bit of work when a simple upgrade should have been functional without all that extra troubleshooting. Glad to see it saturating all the vRAM of my 4090 without overflowing. It also doesn’t seem to be respecting my 3840x width requirement, and I’m getting of 5460x8192 outputs for batch processing. Not a big deal as I can easily fix that in Bridge, just thought it was interesting. Upscaling my FB archive from 2007-2020 to throw into my Immich library.
And does wonder v2 work locally for you now?
Nope, v2 has the same issue I had with v3 before doing a full clean install. Might work if I tried the same, but I’m focusing on a very long project, and v3 seems to be working.
There is a known issue that we had a model bug in V3. We are releasing a new version tonight that will fix it. Thanks for the feedback.
Same problem here ![]()
Bug Report: Permanent “Analyzing Photo” Loop on High-Thread Workstation
System Environment:
CPU: 64-thread Workstation (Threadripper/Ryzen)
GPU: NVIDIA GeForce RTX 4080 SUPER
OS: Windows 11
App Version: Topaz Photo v1.5.0 (Clean Install via Revo)
The Issue: Topaz Photo enters an infinite “Analyzing Photo” state. Even when processing small individual cutouts (single face), the UI remains stuck on the analysis animation for 10+ minutes.
Log Analysis Findings:
GPU Rejection: The tzlog shows that for core models (specifically sms_new, txtdtx, and gfp-l), the app fails to find valid GPU kernels and defaults to “OX with CPU”, despite the RTX 4080 Super being manually selected in Preferences.
Thread Bottleneck: The engine attempts to spawn 32-64 parallel instances of the AI models to utilize the workstation’s core count. This causes a massive delay in model initialization (taking over 15 seconds just to load a single face-recovery node).
Authentication/Model Fetch Error: The log reports multiple 401 Unauthorized errors when attempting to fetch model-specific weights for the current GPU, preventing the high-performance kernels from loading.
Face Count Overhead: On the original class photo, the Autopilot logic hangs when attempting to generate masks for 53 faces simultaneously on the CPU fallback path.
Steps Taken:
Performed a clean uninstall/reinstall using Revo Uninstaller.
Verified “Download All Models” was completed during installation.
Manually locked AI Processor to RTX 4080 SUPER.
Toggled Face Detection to “None” in Autopilot settings.
Wonder 3 is definitely worse. Seems like it’s applying denoise and trying to add HDR to the image…
Yes, it definitely worse than Wonder V2!
The Faces look like Plastic, some nice little Details get oversmoothed etc
Do you have logs from the sessions where the analysis was taking a very long time? Will need those to identify the exact cause as we have not yet reproduced this issue.
Does it happen on every session or only the first session?