Inconsistent RF-DETR Inference Results Between Roboflow UI and Local Inference (Preprocessing and Post-processing Clarification)

Search before asking

• I have searched the Inference issues and found no similar feature requests.

Question

Hi Roboflow Team,

I’m using an RF-DETR model trained on Roboflow and observed inconsistent inference results between the Roboflow UI and my local inference setup.

Specifically:

When running inference via the Roboflow UI, detections are accurate and as expected.
However, when I run inference locally (using the HTTP API and other methods), the same images produce noticeably different results (missed detections / shifted boxes / confidence variations).

I suspect this might be due to differences in preprocessing (e.g., resizing, normalization, color handling) that are internally handled in the Roboflow pipeline but not clearly exposed for local replication.

Could you please clarify:

What exact preprocessing steps are applied during RF-DETR inference in the Roboflow UI?
Are these same steps automatically applied in the hosted inference API, or do we need to manually replicate them for local/custom inference?
Is there any documentation or reference implementation to ensure preprocessing consistency?

Understanding this would help ensure consistent results across environments and avoid performance discrepancies.

Thanks in advance!

Additional

No response

No file, test, or entry point is named in the issue. Start by locating the RF-DETR inference preprocessing and post-processing paths, then compare them with the Roboflow UI and hosted HTTP API behavior; done means documenting the exact steps and whether local callers must reproduce any of them.