
Cute Witch tea party
Image with prompt : https://novelai.net/explore/image/958b3220-2d9c-491a-b295-88e8a477a081

Image with prompt : https://novelai.net/explore/image/958b3220-2d9c-491a-b295-88e8a477a081
I don't know how to really explain it, but my girls always have a slight bulge in the bellybutton area.
This is a picture with "stomach bulge" at -5, "big belly" at -2 and also "pregnant, fat, belly, big belly" at 5 in the undesired section.
Most annoying part is, sometimes I don't get the bulge with the exact same prompt. If I get a picture of her sitting upright, it's not there, but if she's lying, it is. And it's especially visible if I put "from below" instead of "from above".
It happened with multiple different artist combinations, however, I do have 3 that I use pretty much always cause I like them a lot. Houkago shounen, taotao and dombear. Maybe they just draw stomachs that way? This wasn't really a thing on V4.5 btw, even though I was using those same artists.
Would appreciate some advice, the first pic has metadata.
There is so little control over the art style in v5 that it feels impossible to use to generate a consistent set of images. Every generation feels more like a gachapon than before and it doesn't handle artist tags as well as 4.5. yet, it is undeniable that v5 handles prompts much better. The tradeoff between art style and prompting feels bad. Really hoping for vibe transfer to be released soon, don't think I'll resub until that's out.
I can generate a couple of characters that weren't on the previous database, some new ones still don't work. But there's this issue where characters from old animes that would generate just fine on 4.5 show up as generic ones on 5. The first image is v4.5, the second is v5, and the third is the original. It's a little annoying to deal with on v5
Hi everyone!
I'm not completely new to NovelAI. I used it quite a lot back when V3 was the only version available, but eventually stopped using it for a while.
Back then, I remember artist tags blending together really well. Now that I'm back, I've been trying to achieve the same results with V5, but it seems almost impossible. Someone recommended trying V4.5 instead, but I'm still having trouble getting it to work.
Am I doing something wrong, or has the way artist tags work changed between versions?
Any tips or advice would be greatly appreciated!
Thank you so much! 😊
2x2 4koma read right to left. ik it's usually top-to-bottom but i found the best stability with this format
came up with a bunch of stupid gags to test V5's comic/manga capabilities
nlp and text rendering vastly improved compared to V4.5
struggles with speech bubble placement a lot but creating a separate character slot for the worst case scenarios tends to help wrangle that
overall V5 looks great for this kinda stuff though
Did they go up? I remember the last time I could get 10k for 11 dollars, now it shows $14 :/
the eastern style is still relatively easy to generate, but more inpaint is required in this batch compared to the last.
the default art style of V5 model is surprisingly consistent comparing to V4.5, though specific prompts might affect that unexpectedly, it is not particularly hard to spot and remove said prompts.
For example if you have 4 elements you like: day, night, castle, sea, you will have a checklist with day - castle, night - castle, day - castle - sea, etc.