marklr / labelizer

generates captions for local or photoprism-hosted photos
MIT License
7 stars 0 forks source link

Feature Request: #1

Closed iluvatyr closed 11 months ago

iluvatyr commented 12 months ago

hey, this looks very nice and promising.

Could you maybe give some data about how long the average library (10000-50000 pictures) would take to process with low-/mid-/high-end cpu/gpu?

marklr commented 11 months ago

On a CPU, Intel(R) Xeon(R) CPU E3-1275 v5 @ 3.60GHz, it takes about 5 seconds per image as things currently stand (2 models, 1 caption 1 vqa + about 10 prompts). Average image is a 5MB jpeg.

on a mac with mps it's about half that. I haven't yet tried on a CUDA-capable device but the code should definitely run. I've also added a basic streamlit runner to enable rapid prompt prototyping in the latest commits.