FEATURE REQUEST: "Refine Image"

#10725
by deleted - opened
deleted

I have no idea if that would work, but im my opinion it would be nice to have something like that...

For example: I write something and Craiyon normally generate 9 images for me.. so i pick one of the 9 images, choose some kind of "similarity factor" to 60% and click on "Refine Image". So Craiyon will generate the 9 images again, but based on the image i choose to refine, with 60% of similarity from original ...

Sounds crazy, and im probably dreaming way too high (as always), but i think it could be useful..

hello perhaps better to ask somme other ai who already do this kind of things if we can importe a pict to do it

deleted

Oh well, i was thinking like.. when you prompt "realistic painting", it shows a lot of realistic paintings, but almost all of them uses different color styles, different traces, etc.. so you use the "refine image" to generate minor variants of the image you like.

I generated 2 pics of girl faces that i love it, and both of them was from different generations and has something in common: a green circle in a white background. But if you prompt that, the AI generates something completely different from what you are expecting. That means the AI used similar images as base to generate both of images, and would be helpful if you sightly change that to generate similar images with a few differences between each other..
So if you refine the image with 80% of similarity, the AI will maintain the 80% of base images and will change the others to give other "perspectives"..
But anyway, i have no idea of how it works, im just trying to explain in the most pratical way possible, without too many programming theories or rules..

Stable Diffusion works nicely for upscaling craiyon images. Here's an example, Carnage vs Victor Zsasz

craiyon_163655_Victor_Zsasz_versus_Carnage__comic_book_cover_br_.png

Stable Diffusion upscales:

canrage-vs-victor-zsasz-marvel-dc-comic-cover-art-high-definition-hd-stunning-awesome-amazin-123358244.png
canrage-vs-victor-zsasz-marvel-dc-comic-cover-art-high-definition-hd-stunning-awesome-amazin-929643437(1).png
canrage-vs-victor-zsasz-marvel-dc-comic-cover-art-high-definition-hd-stunning-awesome-amazin-929643437.png

Carnage vs Lobo (craiyon, original)

craiyon_124737_Lobo_versus_Carnage_br_2.png

Upscales

--683581859(3).png
--683581859(2).png
--683581859(1).png
--683581859.png

Mess about a bit with the settings. In this I did either just a space for the prompt, or simply 'hd' just to keep it purely based on the image. No filter. Image strength from around 30 to let it be a bit more creative, up to 60 to recreate the image more faithfully (but it will be a bit more blurry because of craiyons quality). Stick with a square image dimension ir it will get stretched out. Prompt guidance can be default (7) unless you want to add more in the prompt. Max quality is 50, so no reason to put it any less I don't think.

Yes, this would be cool

CraiyonTeam should add a generate variations button like dalle has.

I worry that Craiyon/DALL-E Mini is crippled intentionally to still make money with a full featured version.

Sign up or log in to comment