Anyone having trouble with the model? I keep getting slammed with "We're sorry — we've run into an issue." errors.

>>2371 Their servers are getting hammered pretty hard. I started getting more and more delays this afternoon, sometimes it would take 15 minutes to output an image, so that's probably it.

Jesus christ bing really is anal about what it wants to censor. You put "belly" anywhere into a prompt and it's going to block it, try and replace it with stomach or midsection and it's equally as likely to get blocked. Just got second suspension because there's no fucking way to figure out what isn't allowed without trying the prompt and variations thereof, now instead of a 1 hour suspension it's 24 hours, a lot of blocked prompts now erroniously appear as an error with their servers being overloaded. I imagine the next time the account will get banned entirely. This is all after getting it to generate numerous images featuring tits with nipples, so it obviously has the data to generate flat out porn if it wasn't so cucked.

>>2373 Well, I tried inputting "The Sea" and I got errors for it. I have no clue how people are even managing to get one request in. I guess we'll just have to wait for another leak like with all the major models.

>>2374 There was only one leak and that was the NovelAI leak. Dall-E has never been leaked. Now NovelAI is back to focusing on language models and the leaker got the paddle, so it's unlikely we'll be saved by another leak. To make matters worse, all the anti-AI cucks are trying to take down Stability AI and ignoring closed tech kike companies like OpenAI and Microshaft. The future is looking grim for freedom in AI, we need the AI equivalent of Stallman.

https://civitai.com/models/156188/penny-inspector-gadget-1983-character-lora Oh boy!

>>2376 Sweet western loli.

Seriously, how the fuck are you getting such good results? It always misses the mark for me unless it's a very generic prompt.

>>2379 also I vote for making a new thread on /v/ for AI images - except with the demand that it MUST be vidya-related. No tangent shit.

Dumping some images I made today based off of Dore's works. Zelda and Elden Ring.

>>2380 >also I vote for making a new thread on /v/ for AI images - except with the demand that it MUST be vidya-related. No tangent shit. I wouldn't mind this at all. Again, if the thread hits the bump limit, remember that we're archiving these on >>>/ais/

The guy doing a lot of AI generation on /loli/ is quitting for the time being. >>>/loli/5991

>>2383 >SDXL It seems SD 1.5 it is still the best there is.

>>2384 It's super disappointing. It had so much promise at first, but now everyone is trying to crack down on it.

>>2385 Bing using the dall-e 3 version is probably the best there is right now, we can only hope for dall-e 3 to get leaked sometime soon. It's damn good, but it's killing me that there's just no actual information on WHAT isn't allowed in a prompt, and other forums have users complaining about the same thing where they've tried innocuous prompts that kept getting blocked, and they systematically removed every word in their prompt, getting blocked every time until getting suspended. So at least I know I'm not the only one pissed that there's no way of knowing what is blocked and getting penalized for trying to figure it out. I think they've banned me now because I ran out of credits, and upon trying another prompt the page has been stuck at 30 second wait time for hours with an error stating they're working on the problem.

All pantsu generation comes out low leg and skimpy. What keywords to give girls more modest pantsu?

Dall-e is a pretty good improvement compared to SDXL. It can actually write legible text and getting different characters in the same image is pretty impressive. It's not without it's fair share of flaws though: >doesn't know who Raziel, Kain or Moebius are >doesn't know Kat or Harry Mason either >hands can still get messy >same goes for legs >no offensive content or ban >no bad words or ban >nobody knows which words are bad, so have fun not getting your image because of whatever >feed microsoft some more data It's pretty fun though, since you can easily have several different characters in the same image it opens a lot of possibilities, but microsoft filtering and censoring things is just annoying. Hopefully this gets leaked and gives the community one hell of a boost.

>>2383 I wouldn't say I'm quitting per se, it's more like recharging my batteries and focusing on IRL work that I need to do and has been piling up. If by the time I'm finished with work there's still no uncucked new model to use, I can still paint by hand and enhance the pics with the old models, it's just going to take a lot of work compared to using better models.

>>2389 I'm not a lollipop person, but would you mind sharing models/LoRas and whatnot so that other anons can pick up where you left off? It's so frustrating whenever anyone puts a lot of effort into something and once he calls it quits no one can substitute them. >>2386 >Bing using the dall-e 3 version is probably the best there is right now, we can only hope for dall-e 3 to get leaked sometime soon. >>2385 According to some cuckchan anons, the only main difference between Dall-e iterations seems to be the number of parameters. The larger the set of parameters the AI has to control, the more complex and more accurate the results. This has zero sources behind it and I suspect that at some point there's a ceiling to how many parameters you can feed a neural network before it overfits, but the results do speak for themselves. >>2388 Since you seem to be doing some prompts, do you think you can get anything related to Dragon Quest on these models? I have seen on Pixiv plenty of artists making AI pictures with those characters, but whenever I try they always come either super generic and off model or the various models do not understand what a Dragon Quest is. Are there like specific LoRas for that? How are the LoRas themselves made, too, are they just trained over and over on a small dataset of pictures?

>>2390 I left a lot of info about my setup and process in the /loli/ thread. But to summarize. The main models I was using were Based64mix, cardosAnimated, RevAnimated and, more recently, the lolidiffusion versions of them. The lolidiffusion models are here: https://huggingface.co/JosefJilek/loliDiffusion/tree/main The others can be easily found on civitai and huggingface. As for LoRas, I was using the detail tweaker LoRa a lot, and also several art style LoRas that I mixed and matched depending on the pic, like 90's vibes, BlueTheBone and Toriyama style. I was also using the easynegatvie LoRa in my negative prompts. The main thing with my pics is that I was using inpainting a lot to fix my pics after the initial gen and upscaling and I left some inpainting and upscaling tips in the /loli/ thread.

>>2391 What kind of specs did your computer have to run those prompts and did you have long waiting times?

>>2392 I was using an RTX 2070. Generating the base pic took like 5-10 seconds at 512x768 with 28 steps and the DPM sampler I was using, upscaling took about a minute.

>>2390 I'd love to set up a dedicated server farm to run stable diffusion it's getting more and more difficult to get any prompts through on this other shit but I've had to spend so much more time learning how to train it for anything, I probably don't have enough image data to feed into it and then I'd have to spend time figuring out how to web crawl for images from all kinds of sources and I'm sure google is unreliable for that these days. If I had time and money to invest into such an undertaking I'd be all over it.

>>2394 also it seems like bing is just breaking constantly, maybe because I've been in the same session for days, none of the blocked prompts are telling me I've violated policy, it just says they're under too much load to generate, I also cannot enter new prompts now unless I refresh the page or cycle back to an older prompt, this includes when the prompt is blocked, and instead of allowing prompts with timed intervals after credits run out it produces an error "the team is working hard to fix this" preventing use entirely

>>2390 I tried to log in again and turns out the account I used yesterday got blocked and microsoft demands a phone number to unblock it, so I had to create another account instead. The same will probably happen to this one. I prompted some simple stuff, but keep in mind that I never played a single Dragon Quest game, so I'm not the best person for this so the images are all very generic. Dall-e doesn't seem to know very well what a dragon quest really is as you pointed out and the quality is not the best. I don't remember the exact prompt, but it was something like "dragon quest warrior holding sword in an open field, sorceress and archer in background, clear blue sky, sunshine, art by akira toriyama", then I added/removed stuff like "monster/slime" or "celshaded/hand drawn". I also tried "dark fortress, dark sky, thunder, storm", but that was basically it. >Are there like specific LoRas for that I just did a quick search on civitai and there are several LoRas for DQ characters: https://civitai.com/search/models?sortBy=models_v2&query=dragon%20quest There is also the Akira Toriyama style LoRa: https://civitai.com/models/4857 >How are the LoRas themselves made I never made a LoRa since I've been lucky enough to find LoRas of a lot of characters I like, except for Dredd and maybe another character or two. Try checking this guide: https://civitai.com/models/22530 Also discovered a few more annoying things about Dall-e: >Hatsune Miku hugging Judge Dredd is offensive I know hugging a judge is an offense and Dredd would fine/arrest Miku, but this is pretty stupid. I'm pretty sure this is what got my account blocked yesterday since it was the last thing I prompted and I tried it twice and again today with the new account. >Hatsune Miku is offensive for some reason The prompt was literally just "Hatsune Miku". I don't get it. >doesn't know LadyDevimon or Dark Magician Girl >smug (woman) is bad >evil (woman) is bad too, but only sometimes, maybe I also tried to prompt that smug blue girl and got a bunch of warnings just because of the "smug evil" words. I didn't dare add stuff like "offshoulder" or "large breasts" or the much dreaded "hugging" word, because I was already expecting to get warned. I got 2 images out of it and was actually pretty surprised that Dall-e generated them, but they really look nothing like that DQ character. I'll post them in the next post.

>>2390 I also prompted some stuff in SD with the Toriyama LoRa out of curiosity. The first 2 images are the smug blue girl form Dall-e, the rest is SD. I was pretty surprised I got these 2, since she is naked and all. I never used the Toriyama LoRa much and most images were terrible, but its much easier to get some DQ with this than on Dall-e, since the LoRa seems to take care of almost everything. I've also run out of good boy points on bing, so that's all I'm prompting on it for today and I assume that my current account will be blocked by tomorrow anyway. >>2395 >seems like bing is just breaking constantly True. It can block you for no reason, the warnings are mostly generic "oops, error! have a blowfish image" or something else that tells you nothing of worth, "sorry, the team is working on it" warnings, "wait 30s" actually meaning hours or who knows how long, etc. I also don't understand why it seems to generate 4,3,2 or just 1 image at times, but it's probably due to load or something. The points get spent all the same, no matter if you got 4 or just 1 image though. Also, if the prompt button stops working, simply try refreshing the page.

>>2397 >why it seems to generate 4,3,2 or just 1 image This is some hearsay I've read from other sources: >Your microsoft account has a 'ban score' >words or names that directly related to violations of the TOS like celebrity names, obscenities, violence, etc are worth the most ban points >every rejected prompt adds to this 'ban score' but does not directly violate topics outlines in the TOS, the content violation being the quickest way to build up this score, this seems entirely based on arbitrarily banned words or context prediction determined by bing's in-house language model, there is no way to know what it will find offensive now or in the future >the second factor 'nsfw detector' being run on generated images will also add to this score, even though the image may not be pornographic or obscene with no way to verify that it actually is, even though it's the AI at fault for making it potentially obscene >when you get less than 4 images, that means that the omitted of the images in the set were flagged by the nsfw detection, and those images still count towards this ban score, again not being the fault of the user and potentially a false flag So it's extremely easy to get banned based on these metrics if they're accurate.

>>2398 also bing is no longer handing out 100 boosts a day, I got 100 boosts the second day I was using it even after a long suspension, now it's only 25 boosts per day, and others have confirmed they're having serious issues with the prompt ui not working, you are supposed to be capable of using it without boosts, just with waiting time, but now once it hits zero you are done for the entire day, the generate button no longer works until refreshing the page, and it is mistakenly telling users their blocked prompt just could not be created due to high traffic more and more users are getting pissed users are also reporting they cannot even utilize microsoft's reward system to get more boost credits, it literally takes them in circles with no way to redeem, some are desparate enough to buy boost for their new addiction but found it's not even an option what an absolute dumpster fire

>>2398 >>2399 >it's extremely easy to get banned based on these metrics A few months ago I tried using bing to create some prompts to then use on SD. One of them was something like "embarrassed woman struggling to eat hamburguer". Bing would not generate it claiming that it was against the rules or offensive, but when I said I wanted a man eating the burguer instead it generated a full story of a guy eating a burguer and choking to death with it. >serious issues with the prompt ui not working I had to refresh the page for every prompt of the 25 boosts from my new account, because the create button would stop working and when it got to 0 boosts it got even worse since it wouldn't even update the page with the images anymore, plus if the images were not ready yet, it would keep increasing the wait time. It's better to just create another account and keep going, at least until microsoft catches on, which should be pretty soon. >some are desparate enough to buy boost for their new addiction This stuff is pretty addictive, but paying for this censored mess is pretty stupid. And microsoft keeps censoring more and more words. I tried to prompt Miku giving Dredd a hug yet again and nothing, a hug is simply offensive. Dall-e also thinks Miku is a group of people sitting around a campfire and the Balrog is a big generic demon. I also prompted some forest nymphs covered in leaves and it was fine, until I specified an "anime style", then I got an instant warning. You can still get some pretty cool stuff out of this since it actually understands concepts much better than SD, but the constant censorship over everything even remotely offensive is very annoying.

>>2400 Yeah, I found out quickly that just "anime" is a no-no word. You can specify manga style and it works fine.

>>2400 >Trying to make sex scenes What was his problem? I guess I'm not interested to generate sex scenes myself, but I've seen other prompters are doing just fine.

So, I need your opinion, mates: do you want the thread to keep going until it reaches the maximum post limit, or would you rather have another one up? I ask you this cause I need to report it to get it transferred to the other board.

I'd personally like a new thread.

>>2403 Make a new one

>>2384 Hi please help me