For all you AI'ers... I have noticed that Qwen does not have very strict nudity checks or any real checks at all going on. I'm sure you all know how annoying these are. Most have a secondary AI model that will examine all your images for 'policy violations' before displaying the result. Super annoying after you waited all that time for an image and it returns nothing.

Qwen is very capable image generator but not half as good as Nano Banana for producing WAM. But if you take an image from Nano and ask Qwen to remove an article of clothing it does it. Sometimes if you specify key words like 'panties' or 'breasts' then they will tell you that violates their policy even before trying to generate an image so you have to be careful with that. But it is fine removing shirts or whatever. It replaces the shirt with whatever it thinks is appropriate. Sometimes it gives just a 'ken doll' set of privates where there is no detail whatsoever in the key region. Or if the girl has been wammed, then it will just put mess there.

Either way good stuff to know. I suspect the Chinese had copied the Google flash image create model which had this effect as a 'bug' until Google fixed theirs.

Also.. useful site called LM Arena allows you to test out multiple AI models without creating an account which is nice.

That's good to know, thanks.

Qwen used to have (IMO) the best video generator of all, and then after an update they completely broke it. I tried it again recently and it was still broken, even if you choose earlier models.

QWEN still is the best open source image generator. The mistake you're making is assuming they're going to train on porn and fetishes and also allow you to generate said images. The same can be applied to nanobanana and the new nanobanana model that will release very soon.

That said, if you can run locally it does very well.

I think for now it is more fun to create AI images than to actually believe/expect them to look real. Nothing quite yet replaces the real thing.

I'm surprised how many topless/nude results I've gotten on Gemini. The thing is... I wasn't asking for toplessness/nudity. I guess that the wording of my prompt allowed the results to be seen a certain way.

And how many nude or topless pictures are of any decent quality. 2 years ago, you could prompt and work around Dalle-3 and produce the same level of nudity. Every "nude" pic I've seen from gemini, looks no better than a mannequin or barbie. What is even the point at that level.. Unless you like dolls of course..

There were nipples in my end result. I was surprised how many times it happened.

Agree QWEN is the best "open source" generator. I would like to be able to run a copy of QWEN but I have a very average video card and I can only run the smaller versions of it which produce poor results. To buy a high end graphics card for me is not really worth it because obviously the technology will get better in the years to come.

And the creators of AI quite frankly are missing the boat when it comes to not embracing the porn or fetish aspect. I mean that is seriously how the internet was created with over 50% of the USENET groups were in the notorious "ALT.Binaries" channels which was nothing but sharing porn images. (Heck I can remember when Rob Blaine and MK put together the first WAM channel and why they had to differentiate by inventing the word wetlook so they didn't get confused with the piss and shit fans out there). AI could really disrupt the porn world and seems like all they want to do with it is disrupt the advertising and entertainment spaces. We'll see how the future unfolds.

A few have said recently but you don't have to run it on your own PC. It's reasonably cheap to rent a server and spin it up to experiment or train. You're never going to get the same quality out of a commercial model whether that's because of guardrails or pure lack of training content. Wam is niche enough that most models aren't trained on decent dataset of the fetish.

As for commercial models embracing porn.. I've no doubt that VEO, Sora2 etc are trained on vast amounts of NSFW content. Unrestricted Dalle3 models existed with little to no guardrails 2 years ago and it was quite possible to generate filth. The same will apply to these modern models.
Why they don't... It's a huge lawsuit waiting to happen with imageedit models and first frame/last frame etc. Also financially, we're not the real target audience of these models. OpenAI and google are courting hollywood and looking to replace traditional film making and VFX studios.

I have never thought of renting a server. That is a great idea. I usually just clone the open source model from GitHub or wherever it is posted and run it locally. Thanks!

Massed Compute or runpod are probably the best options. Massed Compute have plenty of servers and generally cheaper than Runpod. But.. you need reset up each time you rent as they don't have paid storage you can fire up.
Runpod have more options with images and storage. But a little more expensive. I've used both for training when I didn't want to tie up my gpu.