Spaces:

jbilcke-hf

/

ai-comic-factory

like 6.94k

Running on CPU Upgrade

This might be something of a tall order here but it represents something I don't think I've seen in any text to image implementation.
If someone is aware of this happening in the wild I'd be very interested in knowing.

The Problem
Text to Image models generate excellent imagery but they do not do well at assigning transparency to areas of the image.
This also extends to a large degree on assigning large areas of solid color which can be removed later.
These solid areas (whether of color or transparency) are important for the use of compositing.

A Little Background for Context
All of the red, green, blue channels of an RBG image are in fact levels of gray (from black to white)... on RGB scale from 0 (black) to 255 (white).
This designates the amount of red, green and blue to display in each of those channels.
The addition of the Alpha channel in RGBA images allows for similar greyscale assignment but this time to specify level of transparency; 0 (black/completely transparent, 255 (white/completely opaque).
Image formats such as PNG and TIF have support for Alpha channels
Formats such as GIF are RGB and do not have an Alpha channel but rather rely on keying out aa target color (for instance hot pink) with that target color ideally not being found anywhere else in the image to prevent inadvertent transparency.

Feature Request for Comic Factory
Having the ability to achieve solid color for masking purposes would be most useful.
This especially for designating either areas internally or externally to a generated image
This to allow easily targeting that area for masking purposes and compositing with other images
If this area is designated in the Alpha Channel of RGBA image even better but the solid color option is the simple form of the requirement.
This would also allow text prompts to potentially composite one image over another to maximize continuity (i.e. maintain one element but exchange another)

Largely unrelated aside: Some of this relates to Four Color theory but that is beyond the scope of this request.

Thanks!

• Rodney

I'll add this image generated with Comic Factory of villains meeting to plan their takeover of the world.
All they apparently lack is alpha channel compositing capability...