Magiv2 on manhwas
any chance the magiv2 would be good/work on manhwa ???
im working on a manhwa Recap tool but the llm character context missrecognize the characters a lot
This model was trained on manga - its structure and so on, so it will probably only be good at that...
Theoretically, MagiV2 and MagiV3 are designed for multi-image processing in the context of entire chapters, which is necessary to maintain narrative consistency.
I'm also working on a program that would work for both manga and manhwa, but matching the characters and turning it into prose is a difficult thing... And no matter what you do for now, the percentage doesn't exceed 90%.
In my humble opinion, the only solution is an omnimodal LLM, preferably fine-tuned where there's also a system that detects characters and classifies them depending on their appearance... But this is all based on probability anyway, and no one has managed to exceed the 90% threshold. As for manhwa, you would have to test it, I haven't tested it with manhwa - you would definitely need to divide it into smaller panels.
If you have any cool repositories or code, I'd be happy to exchange them.
For your application, MagiV3 would be more suitable?
- Generating prose/literary narrative
I use the Gemini API for analysis and narration, and I鈥檝e been getting good results with it. I鈥檓 not a programmer, and I built most of my code with AI assistance, but if you鈥檙e interested, I鈥檇 like to get a second opinion from someone with experience in the field.
This is one example of my results so far