I’m running into an issue where local AI processing in the new Topaz Gigapixel (v1.3.0, still present in v1.3.1) doesn’t actually apply the selected model. Export finishes almost instantly regardless of scale factor, and the output image looks like a plain resize/bicubic upscale — no AI detail enhancement at all.
System: Intel Core i9-14900K, RTX 4070 Ti 12GB (latest Studio driver), 64GB DDR5, Windows 11 Pro 25H2 (build 26200.8655).
What I’m seeing:
Happens with every model I’ve tried with both Auto and manual parameter settings.
Output file is written in a few milliseconds — way too fast for any real AI inference.
Cloud processing with the exact same models/images works fine.
The older Gigapixel AI v8.4.4 (installed alongside) processes the same images locally without issues, including the generative models — so it seems specific to the new v1.x local engine on this setup.
What I’ve already ruled out:
Model files are present and intact (58GB / 486 files in the models folder)
AI Processor set to “Auto” in Preferences
Repaired the install, did a full GPU driver DDU + reinstall (Studio), updated motherboard BIOS to latest
Tried both Polish and English Windows regional formats (this fixed part of an unrelated decimal-separator issue I noticed in the logs, but not the main problem)
If anyone wants to check their own .tzlog files during a local export, look for a line like:
[AIE] Can’t decode settings JSON <…long base64 string…>: [json.exception.type_error.302] type must be number, but is array
immediately followed by:
Process round is for pre-img-optimization
In my logs this happens on every single local export, right after the “AutoParam” models (ggn_ap/gde_ap) calculate processing parameters — the result fails to decode as JSON, and the pipeline falls back to a plain resize pass instead of running the actual enhance model.
Has anyone else seen this, especially on a hybrid P+E-core Intel CPU (12th/13th/14th gen) with an RTX GPU? Any workaround would be hugely appreciated — for now I’m stuck using the old v8.4.4 or cloud processing.
Update: just tested with “High Fidelity” (a standard, non-generative model) at default settings, and got the exact same result — export in ~2ms, output is an unenhanced resize, and the log shows the identical signature:
[AIE] Can’t decode settings JSON <…>: [json.exception.type_error.302] type must be number, but is array
Process round is for pre-img-optimization
So this isn’t limited to the generative models (RecoveryXL/Redefine) — it affects local processing across the board, including the standard enhance models like High Fidelity. The “AutoParam” calculation (ggn_ap/gde_ap) runs before processing regardless of which model is selected, fails to decode its result every time, and the pipeline falls back to a plain resize. So if you’re on v1.3.x and any local export feels instant / looks unenhanced, it’s worth checking your .tzlog for this exact error — doesn’t matter which model you picked.
Can you share the original image and the enhanced one? We have a machine with a RTX card that we can test.
We also have a ticket where we track similar cases, but so far it’s very few people and no common link between the cases except we see that line in the logs. I added you there.
After a lot of troubleshooting (and great help from the support team), the issue is fixed. Posting the full solution here in case anyone else runs into the same thing.
Most possible cause: a conflict between the old Gigapixel AI v8.4.4 and the new Topaz Gigapixel v1.x. Having both installed at the same time left behind files, registry entries, or cached model data that corrupted the AutoParam pipeline in v1.x — causing the “Can’t decode settings JSON: type must be number, but is array” error on every local export, regardless of model or scale.
What actually fixed it:
Used Revo Uninstaller (free version, Advanced mode) to remove BOTH Topaz Gigapixel v1.3.1 AND the old Gigapixel AI v8.4.4 — this is the key step the standard uninstaller misses.
Ran full Windows Disk Cleanup.
Fresh install of Topaz Gigapixel from the downloads page.
After that, local processing works perfectly across all models.
One more thing worth noting: I originally suspected the Polish Windows locale (comma as decimal separator) was causing the JSON error, since the logs showed “gamma: 0,750000” with a comma. After the clean reinstall I switched back to Polish regional format — and it works fine with the comma. So the locale was NOT the cause, just a coincidence. The real culprit was the v8.x / v1.x conflict.
If you’re on v1.x and seeing instant exports with no enhancement applied, check your logs for:
[AIE] Can’t decode settings JSON …: type must be number, but is array
Process round is for pre-img-optimization
If that’s what you see — and you previously had Gigapixel AI v8.x installed — try removing both versions with Revo Uninstaller in Advanced mode and doing a clean reinstall of v1.x only.
