I have a few questions regarding this finetune.
Hi, I was wondering the following things regarding this Anima Finetune:
- Was Genbooru or Danbooru used for the updated database? Because I read that Genbooru was mostly used to train Anima, dunno if adding Danbooru only would affect the model.
- Any plans on adding more Western Cartoon series to the data? It would be nice to be able to create images with series as old as Tom and Jerry, Flintstones, Kim Possible, etc.
- Regarding styles, would the model also update more real world knowledge? Like photographic images, realistic style, more world locations, specific races (asian, european, indian, etc)?
- Will the database stay up to July 2026 or are there plans to update it as the model keeps evolving towards 1.0?
- Lastly, you said the model has 40 blocks instead of 28, did you train specific block for specific purposes? It would be nice to have specific blocks just with character data, others with style data, etc. Because the current 28 block Anima has most of its Style and character data in a lot of the same blocks.
Anyways, thank you for your time, I'm looking forward to see this finetune evolve! Thank you for your efforts!
- Both Danbooru and Gelbooru were used, the dissimilarity occurs in less than 0.1% of the images, mainly concerning artist names and some character names (which does changes occasionally even during the collecting data process).
2-3. To certain extent, but I want the model to be an East-Asian style anime illustration focus model.
The dataset will be continuously updated.
All blocks do a bit of everything, but some blocks are experts at certain things (early blocks handle composition/high noise better, late blocks are good at fine-detail). This occurs naturally during training because of the optimizer.
Thank you for the response! I also agree that East-Asian style model is the best due to the amount of content and characters they provide, but would be nice to have some extra knowledge in Disney movies and old classic cartoon styles, which kinda lack a bit on Genbooru/Danbooru. Hopefully its possible, would make the model more creative imo. Same with the realistic styles, it seems Cosmos 2B might not be that good at realism, but having that data I think it might help the model in regard to environment consistency and/or natural body proportions when asked.
Either way, whatever you do will be a improvement to the model, hopefully the new data doesn't affect the stylistic styles from artist or characters/series. Again, thank you for taking time to answer, looking forward to test this model more!