How long does an AI headshot take? We measured it

Key findings

1. The median FaceShot headshot was ready 23.2 seconds after the user pressed generate. Half of all generations took between 21.8 and 24.8 seconds.
2. 65 of 65 generations (100%) finished in under 60 seconds. The slowest took 36.3 seconds, and 95% took 28.7 seconds or less.
3. 62 of 65 (95.4%) were ready in under 30 seconds.
4. Standard and HD, both 2K, take about the same time: the median was 23.1 seconds for standard and 23.5 seconds for HD. Print-ready 4K took 35.6 seconds, but that is only 2 generations, so treat it as indicative.
5. Requests did not queue: the median wait between the request and the start of generation was 0.07 seconds.
6. Not every request succeeds. In the same period 7 of 73 single-style requests (9.6%) failed, and FaceShot refunded the credits for each of them. 4 of the 7 asked for an image model the server no longer accepts.

Time from request to finished headshot

Mean: 23.9 s.

Distribution

By quality

Methodology

• Source: FaceShot's production database, queried read-only on 2026-10-26. Only aggregates left the database; no images, names, emails or account IDs were read out.
• Measured time: from the moment the generation request is stored to the moment the finished headshot is saved and marked complete. Both timestamps are written by the server.
• Included: completed single-headshot generations ("instant" mode) by regular user accounts, one style per request, not a refinement of an earlier result.
• Excluded: our own test accounts and admin accounts, background batch jobs, multi-variation jobs, records whose timestamps were restored from a backup in June 2026 and no longer show the real timing, and 1 completed generation in the period that has no server processing timestamps. Generations before 2026-06-09 have no reliable timing, so the period starts there; 2026-10-18 is the last generation that matched.
• Failed requests are not in the timing figures. They are counted separately: single-style requests by the same accounts from 2026-06-09 to 2026-10-26, with any number of variations: 66 completed and 7 failed.
• Percentiles are continuous (PostgreSQL percentile_cont), rounded to 0.1 seconds.

The query

with sample as (
  select g.quality_tier,
         extract(epoch from g.processing_completed_at - g.created_at) as seconds
  from generations g
  join users u on u.id = g.user_id
  where g.status = 'completed'
    and g.generation_mode = 'instant'
    and g.is_batch = false
    and g.parent_generation_id is null
    and g.processing_time_ms is not null
    and g.processing_started_at <> g.created_at
    and u.role = 'user'
    and u.email !~* '^[a-z]+-[a-z0-9]+@'
)
select count(*),
       percentile_cont(0.5) within group (order by seconds) as median_seconds,
       percentile_cont(0.95) within group (order by seconds) as p95_seconds,
       max(seconds) as max_seconds
from sample;

Limitations

• Small sample: 65 generations from 48 accounts over about four months. Busier periods or a slower upstream image model could change the numbers.
• The time starts when you press generate. Uploading your selfie and choosing a style come before that, and the app checks for the result every 2 seconds, so you may see it up to 2 seconds later.
• About 9.6% of requests failed in this period. A failed request is refunded, but you have to start it again.
• Multi-variation and batch jobs generate the first image straight away and the rest in a background batch, often minutes later. They are excluded here.
• We did not time competitors. Their turnaround figures in our pricing benchmark are their own statements, so this page does not compare speeds.

Questions this data answers