What setup is everyone currently using?

[LawdStalker]

I've recently started experimenting with the translator again, now that I have access to an CUDA capable GPU. I noticed several new features have been added since my last time using the tool.

Right now, my current setup is:

• Text detection: ysgyolo
• OCR: one_ocr
• Inpainting: lama_large_512px
• Translation: A local LLM (quality is still mediocre, but I'm still trying different models)

I've seen some paid services delivering notably decent output, and was wondering if there was a more optimal configuration that produces similar results with minimal post-editing.
Any recommendations on models or pipelines that are working well for you?

[bropines]

• Text detection: custom
• OCR: google_lens_exp
• Inpainting: lama_large_512px
• Translation: Export text -> send to GROK.

I usually translate R18 content, and the filtering of generations in grok is almost zero if you do a couple of things. And I test my own detector, because I'm usually not satisfied with the results of the current ones, I just train my model. Google still has the smartest OCR. If there is no network, I use oneocr.

[LawdStalker]

That google lens OCR option is pretty good, even though it slows the whole process down a lot.
And now that you mentioned your translation method, I hadn’t even realized there was an export/import option😅. "Manually" translating does make more sense though, especially since you can give the translator better context.

I’ve been tweaking the default LLM API prompt a bit to get more explicit output, which has improved the results. Still, every time I run it, I just hope it won’t self-censor and mess up the output.

[bropines]

I just hope it won’t self-censor and mess up the output.

That's why I use the free GROK in the web interface. In fact, I know the method of bypassing Google censorship, but I will not release it, for the sake of the security of both users and the security of the method. Because according to Google's rules, they can demolish the entire account without the right to restore.

[LawdStalker]

but I will not release it

Yeah, that makes sense. If too many people start using the same method, it could draw attention and increase the risk of it getting patched it out.

[Alexamenus]

I have the main program running on a laptop with a customized json module sending translation requests to another computer over lan running the SUGOI14B llm with langchain
CTD usually
i have to choose the font size manually because text intersection on this program is unbearable on automatic.
mit48px
llama large 512

[LawdStalker]

SUGOI14B llm

I heard this model is behind a patreon. Is it any good?

text intersection on this program is unbearable on automatic

Yeah, I’ve noticed it’s a bit better now, but text still overlaps when speech balloons are too narrow.

[Alexamenus]

SUGOI14B llm

I heard this model is behind a patreon. Is it any good?

text intersection on this program is unbearable on automatic

Yeah, I’ve noticed it’s a bit better now, but text still overlaps when speech balloons are too narrow.

Yeah it's probably the best option now if you don't want to pay for anything and have a computer that isn't from a decade ago. It's more than readable, it just has all the same issues everything else does with pronouns in japanese and what not, which I have somewhat corrected by integrating langchain through running the llm with llama.cpp.
Also I gave up on the auto sizing option, I just manually set text size for every project, have to make things tiny so they don't go out of frame of the canvas because there's no border restriction on text boxes for some reason.