Can I paint some colors as a starting point (like img2img) and let it finish the rest? What about a conversation, like, "the dress isn't blue, it should be orange", doing on top of previous prompt?
not yet, these are great suggestions. it's always a dilemma to add features to mitigate the performance of a weak model, instead of making a better model. most of the problems go away with a better language and colorization model, and many model-specific features are made in vain
A better AI model helps a lot with the first goal, but help only so far with the second one. Truth to be told, there is a lot of contextual color information in black and white photos that an AI model can exploit; but nothing beats someone that knows, for sure, the color of the dress of someone in the photo.
I mean, take a look at https://www.reddit.com/r/ColorizedHistory/ - some of those color artists do a lot of research to know the exact shade of green of the military uniform of some country in the 19th century, and things like that, just to have an accurate reference.
So I think that the ability of directing the color output (either by rejecting a color textually, or by painting over the figure with a starting point - even if maybe I'm not painting with the exact tone or texture but a rough color that should help the AI to figure out the details) is essential for a colorization product, even if the model is flawless!