If the scaling is incorrect somewhere and the model has to enlarge an image that is much too small, then that could well be the case.
A description with the help of gemini.
""Subject: Critical Locale Bug - Decimal Separator Mismatch causing Scaling Issues
Issue Description:
There is a significant discrepancy in how scale factors are handled between US and European (specifically German) system locales.
On systems using a comma (,) as the decimal separator, the AI models receive incorrect or truncated input data, which does not occur on US-based systems.
Technical Details:
Root Cause: The UI (Qt/QML) respects the system locale and formats decimal values using a comma.
However, the AI backends (ONNX, TensorRT, etc.) strictly expect the US standard (period/dot).
Truncation Error:
When a value like 0.5 is passed as “0,5”, the backend engine often truncates the string at the comma, interpreting the value as 0. Similarly, a scale of 2.5 is misinterpreted as 2.
Resulting Symptoms:
Tiny Input/Output:
The model processes an extremely small fraction of the image or scales to an incorrect, minimal resolution.
Red Marble Artifacts:
The mathematical overflow caused by truncated scale factors leads to memory addressing errors in the VRAM, resulting in uninitialized data being rendered as red/magenta marbled noise.“”
Could also lead to an offset of an AI model in the Wonder 2 local model.
Wonder 1 consists of three models.
I think this is also the case with Wonder 2.