ai-image-desc

describe image in English with AI

Preview

Images:

images/test1.jpg
images/test2.jpg
images/test3.jpeg

Running

./aiid-describe -g images

Gives

images/test1.jpg → a car driving down a dirt road in the fall
images/test2.jpg → a person standing on top of a sand dune
images/test3.jpeg → a boat is in the water near a city

Setup & Install

Create virtual environment

./setup-env.sh        # nvidia gpu
# or
./setup-env-rocm.sh   # amd gpu

Install

./install.sh
# Creates symlinks to shell
# scripts that run the
# python scripts in a venv.

aiid-describe --help

usage: aiid-describe.py [-h] [-s] [-g] [-b BATCH] [-lb LOAD_BATCH] [-o OUTPUT] input

Describe images in English using BLIP

positional arguments:
  input                 image file or directory with images

options:
  -h, --help            show this help message and exit
  -s, --show            show image with label in title
  -g, --gpu             use gpu
  -b, --batch BATCH     batch size for GPU processing
  -lb, --load_batch LOAD_BATCH
                        batch size for loading images into memory
  -o, --output OUTPUT   output CSV file to save results

aiid-nsfw-detector --help

(modded describe.py to detect nsfw words auto moderation for a web platform)

usage: aiid-nsfw-detector.py [-h] [-g] [-c number] [-a] [-s] [-sc] [-sn] [-pc] [-b number] [-lb number] [-oc clean.csv] [-on nsfw.csv] input

Detect NSFW images using BLIP (auto moderation)

positional arguments:
  input                 image file or directory with images

options:
  -h, --help            show this help message and exit
  -g, --gpu             use GPU
  -c, --count number    how many captions to generate per image (default: 1)
  -a, --all             output all generated captions (default: output only matched caption or last if no match)
  -s, --show            show images (clean+nsfw) with caption in title after processing all
  -sc, --show_clean     show clean images with caption in title after processing all
  -sn, --show_nsfw      show nsfw images with caption in title after processing all
  -pc, --print_clean    print clean images (default: don't print)
  -b, --batch number    batch size for GPU processing
  -lb, --load_batch number
                        batch size for loading images into memory
  -oc, --output_clean clean.csv
                        output clean image(s) with caption to CSV file
  -on, --output_nsfw nsfw.csv
                        output nsfw image(s) with caption CSV file

Example

• use gpu
• generate 15 captions
• loading batch 512 images ~ 20gb ram usage
• save nsfw to nsfw.csv and clean.csv
• show nsfw images

aiid-nsfw-detector -g -c 15 -lb 512 images -on nsfw.csv -oc clean.csv -sn

aiid-group --help

(needs manual tweaking to actually work well)

usage: aiid-group.py [-h] [-s] [-g] directory

group images by eng names

positional arguments:
  directory   dir with images

options:
  -h, --help  show this help message and exit
  -s, --show  show image with label in title
  -g, --gpu   use gpu