An insurance app sends photos to a vision-enabled chat model to read small printed text…

How it works

When you send an image to a vision-enabled chat model, the optional detail parameter in the image_url field controls how the model processes it. It takes one of three values:

• low: The model doesn't turn on "high res" mode. It processes a lower-resolution 512x512 version, which gives quicker responses and uses fewer tokens when fine detail isn't crucial.
• high: The model turns on "high res" mode. It looks at the low-resolution image first and then generates detailed 512x512 segments from the input. Each segment uses double the token budget, which allows a more detailed reading.
• auto (the default): The model chooses between low and high based on the size of the input image.

The docs spell out the trade-off under input limitations. The low-resolution setting gives faster responses and uses fewer input tokens for some use cases, but it could affect the accuracy of object and text recognition within the image. Small printed text on a damaged part is exactly the kind of fine detail low mode can miss. Two other limits are worth knowing: the maximum input image size is 20 MB, and you can send up to 10 images per chat call.

The other options, and when you'd use them

• Low detail only affects the output language: The detail value changes how the model interprets and processes the image, not which language it answers in.
• Low detail disables OCR and returns only captions: Low mode still processes the image, just a lower-resolution 512x512 version, so recognition of fine text gets less reliable rather than switched off.
• Low detail sends the image to a different model family: detail is a setting on the same request to your deployment, and it only changes the resolution mode.

Exam tip

Lower cost or latency and only coarse content means low. Small text, fine detail or accuracy-critical reading means high, or leave auto and accept the extra tokens.

References

• Use vision-enabled chat models