EveryDream/doc/AUTO_CAPTION.md

5.6 KiB

Automatic captioning

Automatic captioning uses Salesforce's BLIP to automatically create a clean sentence structure for captioning input images before training.

By default this requires an Nvidia GPU, but is not terribly intensive work. It should run fine on something like a 1050 Ti 4GB. You can even run this on the CPU by specifying --torch_device cpu as an argument. This will be slower than running on a Nvidia GPU, but will work even on Apple Silicon Macs.

EveryDream trainer no longer requires cropped images. You only need to crop to exclude stuff you don't want trained, or to improve the portion of face close ups in your data. The EveryDream trainer now accepts multiple aspect ratios and can train on them natively.

But if you do wish to crop for other trainers, you can use Birme to crop and resize first. There are various tools out there for this.

Execute

Place input files into the /input folder

python scripts/auto_caption.py

Files will be copied and renamed to the caption as the file name and placed into /output.

Colab notebook

This will run quite well on a T4 instance on Google Colab. Don't waste credits on more powerful GPUs.

https://colab.research.google.com/github/victorchall/EveryDream/blob/main/AutoCaption.ipynb

It should work on other GPU providers on minimal power Nvidia GPU instances, but you are on your own to upload and download files.

Additional command line args:

--img_dir

Changes the default input directory to read for files. Default is /input

python scripts/auto_caption.py --img_dir x:/data/my_cropped_images

--out_dir

Changes the default output directory. Default is /output

python scripts/auto_caption.py --out_dir x:/data/ready_to_train

--format

The default behavior will simply name the file the caption .EXT and, if needed, add _n at the end to avoid collisions, for use with EveryDream trainer or Kane Wallmann's dream booth fork.

ex output: "a man in a blue suit and a woman in a black dress standing next to each other in front of a table with a potted plant on it.jpg"

"mrwho" or "joepenna" will add [number]@ as a prefix for use with MrWho's captioning system (on JoePenna dream both fork) which uses that naming standard to avoid file name collisions.

python scripts/auto_caption.py --format "mrwho"

"txt" or "caption" will create a ".txt" or ".caption" file instead of renaming the image. ".txt" sidecar is another option for EveryDream trainer instead of getting the caption from the filename itself, and ".caption" is an option for other trainers.

python scripts/auto_caption.py --format "txt"

or

python scripts/auto_caption.py --format "caption"

Tweaks

You may find the following setting useful to deal with issues with bad auto-captioning. Start with defaults, and if you have issues with captions that seem inaccurate or reptitious try some of the following settings.

--nucleus

Uses an alternative "nucleus" algorithm instead of the default "beam 16" algorithm. Nucleus produces relatively short captions but reliably absent of repeated words and phrases, comparable to using beam 16 which can be adjusted further but may need more tweaking.

python scripts/auto_caption.py --nucleus

Beam vs Nucleus

See q_factor below. 0.3 to 3 seem to produce sensible prompts, though 0.01 and 2000 will still work fairly well.

Additional caption example for above with nucleus with different q_factor values:

nucleus q_factor 9999: "a number of kites painted in different colors in a ceiling"

nucleus q_factor 200: "a group of people waiting under art hanging from a ceiling"

nucleus q_factor 1: "several people standing around with large colorful umbrellas"

nucleus q_factor 0.01: "people are standing in an open building with colorful paper decorations"

nucleus q_factor 0.00001: (same as above)

--q_factor

An tuning adjustment depending the algorithm used.

For the default beam 16 algorithm it limits the ability of words and phrases to be repeated. Higher value reduces repeated words and phrases. 0.6-1.4 are sensible values for beam 16. Default is 1.0 and works well with the defaulted value min_length of 24. Consider using higher values if you use a min_length higher than 24 with beam 16.

For nucleus (--nucleus), it simply changes the opinion on the prompt and does not impact repeats. Values ranging from 0.01 to 200 seem sensible and default of 1.0 usually works well.

Beam vs Nucleus

--min_length

Adjusts the minimum length of prompt, measured in tokens. Only applies to beam 16. Useful to adjust along with --q_factor to keep it from repeating.

Default is 22. Sensible values are 15 to 30, max is 48. Larger values are much more prone to repeating phrases and should be accompanied by increasing --q_factor to avoid repeats.

python scripts/auto_caption.py --min_length 20

python scripts/auto_caption.py --min_length 34 --q_factor 1.4

Q vs Min for beam

Note

If you continue to both increase min_length and q_factor with default beam algorithm in an attempt to get a really long caption without repeats it will generate oddly specific prompts. For example using the above image:

--q_factor 1.9 --min_length 48:

"a painting of a group of people sitting at a table in a room with red drapes on the walls and gold trimmings on the ceiling, while one person is holding a wine glass in front of the other hand"