The Voynich Ninja

Full Version: Reverse DALL-E
You're currently viewing a stripped down version of our content. View the full version with proper formatting.
Does anyone know of software that will take an image and then create a description of that image in words? (Preferably conforming to some overall format or formula.) Ideally this would be applied to multiple separate drawings of plants; isolating features of each plant and also common features.
ChatGPT and BARD as most common "AI" generators can also read images and produce textual description of it.

In ChatGPT you need the plus subscription to use that feature and while it is nice it is maybe yet more funny than great.
Never tried BARD so cannot comment on that.

Clip/Interrogator is interesting too and it is free. You are not allowed to view links. Register or Login to view.
Below is what it said about the image here

"a book with a drawing of a tree and a lizard, image apothecary, valerian, hydra, floating kelp, byzantine, l - system, noxious, asymmetrical composition, michellin star, tunic, listing image, colored drawing, spritesheet, medical image, depicting a flower, ponds"
Ideally I would something formalised, almost a script.

FLOWERS: NO (OR SMALL OR ONE OR ...)
ROOTS: HAIRY(OR SMOOTH OR ...
LEAVES: SMOOTH BLUE (OR CURLY OR...)

then I can statistically cross compared plant script text with associated EVA text to see if there are relations between the two. I want the script to be as simple as possible without being too simple. So devising an appropriate script format seems quite difficult.

Of  course a script will not be as flexible to perfectly reproduce the original drawing, nor does it need too. It just needs to describe the core and most significant features.

To select the terms used one needs to compare which are the common features that we see amongst the different herbal drawings.
Something like that is pretty easy to do by hand. It will take a number of hours but the result will be many times more reliable than what the current generation of AI can produce.
Colours and Counts might be relevant

RED, GREEN, BROWN, BLUE ...

ONE, TWO, THREE, MANY...

Terms like

BULB, STRAIGHT, CIRCULAR

might be of use

One might want to define a specific common structure with a given name.
Of course this approach may deliver no useful results. It remains one to be explored.
Many herbs have unique features so it may not be possible to match these.