The ComfyUI workflow to run this is provided as a file as well. (its named as Training data in model files and its a few kb in size)
Colorize Black and white Hentai Manga.
As of now this checkpoint is only tested for hentai manga and not for action or other genres.
Please use the prompt
"colorize this manga image."
You can use the basic Kontext-dev inference template from ComfyUI for using this model. In place of the diffusion model, use this checkpoint merge in fp16. I have not tested this extensively with fp8, but it works.
Set Flux guidance scale as 4.0. higher values would lead to more saturated colors, but might look unnatural and lower values make the image look duller. You can play around with it.
Prompt: "colorize this manga image"
steps 15 (10 minimum)
cfg: 1.0
euler
simple
Description
FAQ
Comments (16)
Thank you for your work. Looking forward to testing this out.
This is a first of its kind and it works so well!!! Have you tried modifying colors using the prompt?
I have, and it becomes worse, probably because of me using a single prompt for training. But color changing with prompt for western style comics seems to work
As of now further development of this isnt planned due to time constraints, but im thinking of making this better with better prompts and possibly going to qwen edit
@Civitai_2 Sounds good, thanks for the tips!! Any more suggested prompts that work, in your experience?
@arkhan9 I have not tested this with other prompts, cause this one works fine for me
@Civitai_2 Therefore, it is not allowed to use the prompt words of the flat coloring style to change the style?
@Ovoooooo that seems to be the case as per my obervation, my assumption is that because I have overfit the model on this particular prompt
Hi, I see you set this to be a checkpoint merge? What is merged into it exactly? A manga colorization lora you trained yourself or something else? Why create a checkpoint merge instead of a lora?
A combination of 2 loras i trained myself with different datasets and training configs
Используйте рабочий процесс с картинки удалите лоры замените диффузию на эту модель. Настройки не меняйте! Сначало протестируйте.
Hello, I really liked the model. It is very accurate. Is it open source? If not, would it be possible to make it open source? I’d love to see how it was built and, if possible, help improve it.
i just took some image pairs and trained the model
Could you upload the fp8 here. I don't know how to convert to fp8.
I will see if I can do it, meanwhile you can use the fp16, just choose load in fp8_e4m3fn in ComfyUI while loading the model
I used Grok to instruct me on how to setup this in ComfyUI, the setup only took a few minutes. The results are really good.
Now the bad part is that on my Laptop with a 1660Ti 6GB Vram and 32GB DDR4 15 steps took around 60 minutes, the final result was amazingly good, tried 3 steps that took 15 minutes, the result was ok, clearly for best quality you need at least 15 steps.
I have considered making a smaller distilled model of the same so the generation gets faster on smaller GPUs, although your gpu in partcilar is going to be very slow, in order to be able to colorize a complete doujin or similar of roughly 30 pages, you would have to leave it overnight even if i do manage to make a smaller model.
I am considering a 4B model as its next counterpart, but time has been a constraint these days
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.