This is Alibaba’s offline text-to-image model, and it has excellent prompt comprehension (LLM-based), high image quality, a higher-than-usual base resolution, and is capable of some naughtypics. It is also quite large (20GB+) and slow, especially when you crank up the parameters to render small text. This image took five minutes to render on my high-end gaming PC:
(the main title always came out correct; the subtitle had about a 50% chance that all the letters would be intact)
Okay, I can work with this (2-minute render):
It doesn’t seem to have some of the training issues I ran into with Omnigen2, but while it’s not fully censored, it’s also not a true NSFW model, and has a tendency to cover the naughty bits unless you specify your request precisely. Checking Civitai, there are people working on LoRAs to provide explicit content, but it takes a while to figure out the quirks of new models. Y’know, if that’s what you’re into…
My prompt started with her naked and added lingerie a piece at a time, but the engine insisted on generating bra and panties; if I didn’t mention any lingerie, she was nude. Similarly, almost every attempt at this sisterhood-is-powerful picture generated a bra, no matter how much text I added to the prompt requesting Dem Tiddies:
It also added the girl in the middle, just this once out of two dozen tries, and I think that really put the cherry on top; it was also one of the few attempts that showed Big Sister from head to toe and scaled the other girls proportionally so they didn’t look like they were pasted in from another photo.
Something I noticed with this one is that “blowjob” was almost never spelled correctly, but “blow job” worked every time. This helps explain why words like “Hornblower” and “harem” are usually wrong: they’re statistically unlikely to emerge from the noise.
I haven’t tried Qwen’s editing yet; hopefully it’s functional in SwarmUI (Omnigen2’s isn’t), because their demos promise the ability to do targeted word-edits to existing images. That’d help sort out the rare-word issue.
Oh, and I did manage to get Our Scout Mistress’ top off a few times, by relentlessly padding the prompt with emphasized keywords:
Markdown formatting and simple HTML accepted.
Sometimes you have to double-click to enter text in the form (interaction between Isso and Bootstrap?). Tab is more reliable.