2026-04-20095633_00002_.png (3.75 MB, 2048x2816) I spent yesterday trying Anima. I initially made the mistake of using 'Anima [Official]' and was unimpressed with everything except the natural language prompting, but switching to AnimaYume made most of my complaints go away. From my testing, it seems to have capabilities somewhere between banana v1 and banana v2, which means it blows most other corpo models out of the water (with the restriction of anime-only content). Natural language prompting-wise, my understanding is that it's effectively an enhanced version of the tag merging (eg. "green fishnet bikini") that you can do on Illustrious, and significantly reduces the need for regional prompting outside of manually specifying the location of things that don't impact the rest of the image. It's not perfect at it, with some minor tag bleed or haziness still happening, so regional could be helpful to reduce those situations. I appreciate the improved character tag quality and recency, but more than that the improved perspective and in-image proportion consistency are the real selling points for me. It doesn't seem to be able to do upscaling though, so there is reason to keep chosenMix around. Its general style tendencies are a little bland too, and it's definitely got a bit too much western art in its training data based on some of the outputs I've had. An odd limitation I've found is that if you have it write text or add an image to an object, it doesn't integrate it well into the object but a pass of chosenMix will add appropriate wrinkles. Overall, a mixed bag but I'll probably switch to it for my less simple gens.
>>50867211 My 3060 takes around 70 seconds per Anima gen, rather than 30 seconds per Illustrious gen, but the relative ease of getting a good gen makes up for that. Besides, I also gen while I read (or watch anime, if my GPU can handle it at the same time).
Anonymous 04/20/26(Mon)05:10:35 No.50870723
2026-04-20095633_00002_.png (3.75 MB, 2048x2816)
I spent yesterday trying Anima. I initially made the mistake of using 'Anima [Official]' and was unimpressed with everything except the natural language prompting, but switching to AnimaYume made most of my complaints go away.
From my testing, it seems to have capabilities somewhere between banana v1 and banana v2, which means it blows most other corpo models out of the water (with the restriction of anime-only content). Natural language prompting-wise, my understanding is that it's effectively an enhanced version of the tag merging (eg. "green fishnet bikini") that you can do on Illustrious, and significantly reduces the need for regional prompting outside of manually specifying the location of things that don't impact the rest of the image. It's not perfect at it, with some minor tag bleed or haziness still happening, so regional could be helpful to reduce those situations.
I appreciate the improved character tag quality and recency, but more than that the improved perspective and in-image proportion consistency are the real selling points for me.
It doesn't seem to be able to do upscaling though, so there is reason to keep chosenMix around. Its general style tendencies are a little bland too, and it's definitely got a bit too much western art in its training data based on some of the outputs I've had. An odd limitation I've found is that if you have it write text or add an image to an object, it doesn't integrate it well into the object but a pass of chosenMix will add appropriate wrinkles.
Overall, a mixed bag but I'll probably switch to it for my less simple gens.
>>50867211
My 3060 takes around 70 seconds per Anima gen, rather than 30 seconds per Illustrious gen, but the relative ease of getting a good gen makes up for that. Besides, I also gen while I read (or watch anime, if my GPU can handle it at the same time).