FEAT: Save generation parameters in PNG metadata

[tildebyte]

This is something I added to my (heavily modified) 'CLIP_Guided_Diffusion_HQ_256x256' notebook last year.

Basically, just as $topic - store all of the parameters used to generate an image in the metadata of the image itself.

Example:

AFAICT, GitHub doesn't munge the image in any way, so opening it with something like TweakPNG (no affiliation) will let you easily read all of the metadata.

The code is pretty straightforward, something like this should suffice

# Required
from PIL.PngImagePlugin import PngInfo

# For example (not necessary)
from torchvision.transforms import functional as TF

# For e.g.
global_config = {
    'clip_guidance_scale': 1000,
    'perlin_init': True,
    'perlin_mode': 'gray',
    # etc.
}


# This stores each param as a separate metadata entry, but dumping the
# whole thing into one field would work just as well
def create_png_metadata(cfg):
    metadata = PngInfo()
    for key, value in cfg.items():
        try:
            metadata.add_text(f'AI_{key}', str(value))
        except UnicodeEncodeError:
            pass
    return metadata


this_metadata = create_png_metadata(global_config)
this_metadata.add_text('AI_this_filename', filename)

# Obviously this relies on tensorflow's image writing...
TF.to_pil_image(image).save(str(filename), pnginfo=this_metadata)

My first thought was to open a PR, but for the life of me, I can't find the image writing code... 😆

[hanxiao]

make sense, will add it as an option

[OverStruck]

Indeed it's a nice idea. I was thinking of doing the same thing but casually decided to see if any one was talking about this and found this! :D