[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]

/vt/ - Virtual Youtubers

Search:


View post   

>> No.75194590 [View]
File: 603 KB, 2048x1024, Hanasaki Miyabi vs. Alban Knox.jpg [View same] [iqdb] [saucenao] [google]
75194590

>>75193433
Oh, sorry if I wasn't clear! Niji 4 and 6 both know Salome perfectly, it's Advent that it's very bad at. Prompt for both sides of >>75192702 is just "Nerissa Ravencroft". Niji 4 are the four on the left, and Niji 6 are the four on the right.

>>75193476
Oh, interesting! I wonder if it was around the same time, then. Niji 6 was released January 29 2024, but it only as a very loose grasp on Advent members. So loose that I can only say that it resembles them if you squint really hard, because it's typically just the color scheme and a few elements that carry through.

Another thing I'm curious about - was NovelAI selectively trained on specific VTubers, instead of a blanket of everything on Danbooru? I know the general idea is "more images on Danbooru = better results" but there's a few niche cases that caught me off guard.

I don't know the exact numbers as of the scrape date, but as of now Holostars member Hanasaki Miyabi has 405 images, and Alban Knox has 250 on Danbooru. Yet, Niji 6 can generate Alban, but not Miyabi.

I also can't attribute the niji・journey devs to favoritism in which VTubers it has been trained on, because I'm pretty sure they don't even know what VTubers are. I have never seen a more milquetoast band of normies work on an anime-styled model before, it's so baffling.

(I can't even find a single user gen of "Gawr Gura" on the user-only galleries. It's crazy how far separated Midjourney users are from the average NovelAI user.)

Pictured: Hanasaki Miyabi vs. Alban Knox

Navigation
View posts[+24][+48][+96]