>>85767447>>85823511Kind of surprised this didn't get filtered
Alright, started testing, and first thing I wanted to know if it really is as good at multiple subjects as some anons claimed.
Short answer: It's crazy good. Understanding of at least holo chuubas is insane (up to advent at least, it somewhat knows ReGLOSS, but not as well, which makes sense with a cutoff after 2023)
Long answer: 2 characters basically always worked
https://files.catbox.moe/6jh5bs.pnghttps://files.catbox.moe/ripptg.pnghttps://files.catbox.moe/svl53z.png (Raden and Hajime, which doesn't work quite as well)
for 3 characters it sometimes messed up minor details, like who is wearing which outfit, but overall still looked good if you chose fitting aspect ratios and prompts(I wasn't able to get two girls to dance and one to be the third wheel, though that might just be me sucking at prompting)
https://files.catbox.moe/g4frf7.pnghttps://files.catbox.moe/yivda9.pnghttps://files.catbox.moe/3wjy8p.pngfor 5 characters, it starts having trouble with clones or missing one character, but a bit of trial and error and surprisingly often it worked pretty well, still the occasional misplacement of features, like a halo or a tail, but really impressive.
https://files.catbox.moe/vdfnz7.pnghttps://files.catbox.moe/8g76ff.pnghttps://files.catbox.moe/2on1xg.pngHaven't tested 4 or 6+ yet, but I'd assume it just follow a trend of getting harder and harder.
>>85777559NR, but absolutely sexo. based request!
>>85793789The light reflection/refraction(not sure what it's called) on her body due to the water is very nice and really sells it
>>85804833Sexo!
>>85815353Welcome back! Nice bike.
>Any new tech rolling around, anything interesting in the AI scene happening?The thing I'm most intrigued by was the OmniGen paper going around a few days ago. As far as I understand they just took an LLM, screwed on a VAE and with a bit of magic, they had a vision model with llm that could gen pretty passable images, all in just ~4B parameters.
Of course, this was just a paper, and now someone would need to sit down and actually make it, but this might be the next step for image models if it actually works as well as they claim.
>>85826652That hammer is awesome!
>>85832490nice set!
>describe other botans being greatLLM response in my AI image thread? More like than you would think!
But seriously, happy you liked them.
>4th tail version? For the onsen images in particular I tried both, and the older version gave me the results I liked better. They didn't really differ much though
045:
https://files.catbox.moe/dd369v.pnghttps://files.catbox.moe/2yun9e.png050:
https://files.catbox.moe/rwu1x9.pnghttps://files.catbox.moe/ai0de8.pngSo yeah, I prefer the older version
