Why is the Processing Time Different for Custom Vision and Computer Vision combined for every run?

Rohit Nidimukkala 0 Reputation points
2026-07-21T07:18:55.3166667+00:00

Screenshot 2026-07-21 at 12.16.55 PM

Screenshot 2026-07-21 at 12.29.21 PM

Screenshot 2026-07-21 at 12.17.46 PM

Hi,

I have built a pipeline that involves sending a base64 format image to a Custom Vision model for classifying the image as PASS or FAIL, followed by a Computer Vision model via imageAnalysis_v4 API for extracting text from paper notes present in the image.
The whole process takes around 9 to 12 seconds to process for the first time, after which the timings vary for the same image when I sent it for processing (screenshots attached) even though I run consecutively i.e. ranging from 3 to 5 seconds. Sometimes at night (IST), I have seen the total processing time go down to even 1 to 2 seconds.

May I know why the processing time is different every time by 1 or 2 seconds for the same image?

Thanks.

Azure AI Custom Vision
Azure AI Custom Vision

An Azure artificial intelligence service and end-to-end platform for applying computer vision to specific domains.

0 comments

1 answer

Sort by: Most helpful
Most helpful Newest Oldest
  1. Christos Panagiotidis 3,301 Reputation points
    2026-07-21T09:27:09.1833333+00:00

    Using the image does not make timing fixed because your measurement includes two API calls, base64 handling, and network transfer. The total cannot show which stage caused the 9–12-second first run or later 1–5-second runs.

    Add separate timers around base64 preparation, Custom Vision, and Image Analysis v4. For each call, record UTC time, HTTP status, region, payload size, and the service request ID. Reuse one HTTP client, place the caller and both resources in the same Azure region where possible, and use Custom Vision prediction without storage unless images are required.

    Compare your request rate with the Custom Vision tier limit; throttling returns HTTP 429, not merely a slower successful response. Run enough samples to compare p50, p95, and p99 for each stage. If one Azure call stays slow below quota, give Microsoft Support its request IDs and timestamps. Identical input can produce identical results without guaranteeing fixed end-to-end latency.

    Was this answer helpful?

Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.