Skip to main content

Vision Preprocessing

OpenVINO GenAI normally selects the most performant vision preprocessing implementation supported by the device passed to the pipeline. For Qwen3-Omni, the VISION_PREPROCESS environment variable can select a specific implementation for debugging and accuracy comparisons.

Qwen3-Omni preprocessing modes​

VISION_PREPROCESSStageProcessing path
Unset or emptyStage 2Uses the default OV implementation.
OVStage 2Forces resize, normalization, patch reshape, transpose, and flatten in an OpenVINO graph on the requested device.
OV_REARRANGEStage 1Forces resize and normalization on the host, followed by patch reshape, transpose, and flatten in an OpenVINO graph on the requested device.
CPPStage 0Forces the complete host C++ reference implementation. No preprocessing model is compiled on the requested device.

The OpenVINO modes are device-agnostic. For example, constructing the pipeline for CPU compiles the preprocessing graph on CPU, while constructing it for GPU compiles the same graph on GPU.

The selected implementation is mandatory. Compilation failures are propagated to the caller instead of silently changing the preprocessing path. Unknown values are rejected. Use OV_REARRANGE or CPP explicitly when diagnosing device support or accuracy differences.

Accuracy considerations​

Stage 1 performs only data rearrangement in the OpenVINO graph and is bit-identical to Stage 0. Stage 2 uses OpenVINO BICUBIC_PILLOW resize to match the antialiased Pillow bicubic algorithm used by the host implementation. Its accuracy is comparable to the host path, but small numerical differences can occur around resize boundaries because device implementations can use different floating-point evaluation orders. Use OV_REARRANGE to isolate rearrangement from resize and normalization, or CPP as the complete host reference when investigating an accuracy difference.

Examples​

export VISION_PREPROCESS=OV_REARRANGE

Remove the variable or assign an empty value to restore the default OV implementation.