search

Particle1904 / DatasetHelpers

Dataset Helper program to automatically select, re scale and tag Datasets (composed of image and text) for Machine Learning training.
MIT License
170 stars 9 forks source link

Great Tool! #2

Closed Seantourage closed 1 year ago

Seantourage commented 1 year ago

Looking forward to the next release. Using it to train photo subjects I made a few tweaks to the tagging that you might want to use in the wiki.

Removed tags: These are the tags I remove from each photo.

real life insert, solo, solo focus, female focus, male focus, solo male focus, solo female focus, mature male, mature, female, muscular male, old woman, fat, fat man, bokeh, blurry background, depth of field, simple background, photo background, grey background, white background, black background, purple background, orange background, horror (theme), star (symbol), photo (object), genderswap (mtf), lips, teeth, nose, forehead, freckles, collarbone, armpits, navel, thick eyebrows, brown eyes, blue eyes, green eyes, black eyes, birthmark, facial mark, scar, mole, mole on cheek, mole under eye, mole on neck, mole under mouth, mole above mouth, mole on body, mole on breast, breasts, nipples, small breasts, medium brests, large breasts, very dark skin, dark skinned, dark skin, dark-skinned male, dark-skinned female, wrinkled skin, colored skin

Replaced Tags 1girl, 1boy, realistic, topless male, poster (object), painting (object), bar (place), canvas (object), mouse (computer), painting (medium) with woman, man, photo, topless, poster, painting, bar, canvas, computer mouse, painting

I'm debating also removing "looking at viewer" since it's so typical.

Also using a 0.25 I did an auto caption of 8000+ photos and these were the most common tags. Used this as a reference for the above (see attached)

tag_occurrences_sorted.txt

Particle1904 commented 1 year ago

Hi! How are you doing? I'm glad you find the tool useful. I don't really understand what is the issue. But if you are suggesting to automatically change or remove tags, its not something I want to do since the tools are supposed to work with booru tags and artwork. Although it works with photographs, its a good idea to also have english captions alongside them.

Seantourage commented 1 year ago

Mainly a thanks and comment since you don't have discussions enabled, just wanted to reach out.