When will training code to be released?

[dk000000000]

Hi, I have tested this model with GrayImage(ScreenStyle) and Sketch with Shading
it seems that both will produce less color.
I belive its because in the training data that older manga seems to use less color.
When you apply to newer manga or doujin, over color tone will look very weird.
I would like to have training code to readjust color tone.

it become real bad when doing multiple sections. is it because some of your training data are partially colorized?

compare colorize done by chatgpt, though its more of regeneration than colorization

[dk000000000]

it seems the model can differentiate whether its from colored image turn into scale, then it will try to do full colorization.
if the model detect its from actually original black and white image it will do partially colorization.
I stronly recommend you clean off those partially colored image from your training data.
And try to use actual black and white image for validation.

[dk000000000]

cobra is indeed better
https://github.com/zhuang2002/Cobra

[dk000000000]

cobra still have body color problem, character skin might be treated as background or cloth

[dk000000000]

i think vlm is needed for image/scene understanding, maybe with control net for detailed color specification