You might be happy to see digiKam devs recently learned about your useful tool

[TFWol]

This post is how I learned about it.

https://discuss.kde.org/t/auto-tag-models-are-poor/33651/3

[jabberjabberjabber]

Thanks for the link. I'm interested to see how it turns out.

[uEsker]

I'm very interested in your project and have been using Digikam's "auto-tags" feature over the past few days (manjaro linux). Its OK, but uses YOLO11 or EfficientNet B7 models with very small tag vocabs. It also writes, it seems to me, to XMP.keywords attribute while also wiping out any file xattrs. This means that even though it has Baloo file indexer integration and can sync back and forth with it, there seem to be issues with network files, baloo indexing fields (not clear what it is using) and filemanager/KRunner usage and availability.

What I'd love to see is your project working with KDE filesystem manager (Dolphin) and Baloo - by writing an xattr under user.xdg.tags. This would mean it would be picked up within the system, Gwenview and Recoll search as well, and have a "standard" KDE point of search/integration. The EXIF writes are good too, so Im not proposing that you would remove them but perhaps have a chain of "taggers" (configurable?) that could output various metadata contents and fields. It seems that there are so many different means of writing tags to filesystems that no single such means is "good" "bad" or "best" and being as flexible as possible opens your work up to more users and tools. (maybe other uses as well - e.g. document classification and tagging?)

Thanks for opensourcing!

[jabberjabberjabber]

I sounds to me like you are asking for a plugin system of sorts to separate the 'tagging' from the 'writing to metadata' portions of the process. Am I understanding it right?

[uEsker]

Yes! And (full of good ideas), it would be great to make the analysis part pluggable as well - so e.g. it could use a chosen model to analyse doc/pdf/xls etc.

Probably taking this too far....but its the kind of thing that could be sold to an organisation (never mind a home-labber like me) that wants to do file "summarisation" but locally, so there no risk of confidential/personal/proprietary info ending up in a cloud via invisible data gathering or telemetry etc. Some might even pay for it....

I've downloaded and run now and its great, I love it. In the mean time Ive also written a quick shell script to find images, use exiftool to pull the field/tag Im intersted in (Subject) and then write to the xattr under user namespace. Then need to separately wait or prompt baloo to update its index and off I go to tag heaven.

[jabberjabberjabber]

Would you be willing to share the script that you made?

[uEsker]

Sure, use at your own risk however, cannot say my testing has been extensive, but there has been some. Two scripts, one to pull the exif xpm.Subject another to write user.xdg.tags. Not much to them, inline below.

First is copyToXattrFromXmpSubject.sh

#!/bin/zsh
# see https://unix.stackexchange.com/questions/29724/how-to-properly-collect-an-array-of-lines-in-zsh
#

arr=(jpg JPEG arw ARW nef NEF raw RAW tif TIF tiff TIFF png PNG webp WEBP dng DNG bmp BMP)
n=${#arr[@]}        ## number of elements in array
regex='^.*[.]\('    ## beginning of regex
srch="${1:-.}"      ## path to search (default '.')

for ((i = 0; i < $n; i++)); do  ## loop over each element
    ## if not last, add "${arr[i]}\|" otherwise add "${arr[i]}\)"
    ((i < n - 1)) && regex="$regex${arr[i]}\|" || regex="$regex${arr[i]}\)"
done

regex="$regex\$"    ## add the final '$'

#find and call copyAttr.sh for each file
find "$srch" -type f -regex "$regex" -exec zsh /home/xyz/bin/copyAttr.sh {} \;

Second is copyAttr.sh

#!/bin/zsh
#
# *** copyAttr.sh ***
#
#copy exif Subject tags to user.xdg.tags so Baloo can index and will show as tags in Dolphin

def copyAttr () {

  arr=(${(@f)$(exiftool -s3 -Subject $1)})

  echo "Got tags: $arr"
  #@TODO:do an update instead of set
  setfattr -n user.xdg.tags -v "$arr" "$1"

  echo "Getting new extended attr"
  getfattr -d "$1"
}

copyAttr $1