>Tried SD 1.5, then SDXL like 2 years ago>Thought it was absolute trash and started genning online>All the massive censorship waves start and the quality of all the online models actually gets worse to the point you can look at, for example, Grok videos from a year ago and then today and their quality has gone down so much it looks like some Indian-tier malware site>GPU dies so I replace my 2080 with a 5080>Decide I'll give local genning a try again, and even knowing I'm handicapped at 16gb vram and have to heavily quant shit, I just wanted to see how it looks>Download Krea 2, Minimax 3, and a bunch of loras >The stuff it makes literally looks like real life, as in I can't tell the difference at all between the pics and vids and actual videos from real events>Gen a bunch of loras of ex-girlfriends (purely for my own use, not to spread online or to show anyone. Fuck off glowies) and the images and videos look like stuff I'd find on an old phone or something. You cannot tell at all any of it is fakeWhy the fuck do people still gen on online models for? The difference is like playing some shitty mobile game and then playing some brand new big budget game on PC on ultra.I was going from spending HOURS of my free time having these sites open in another window trying to gen a chick in a bikini or being able to even see cleavage or an ass crack, with hours of "CONTENT MODERATED" shit, no matter the site used, and now I can just gen completely life-like shit in seconds that can have full text adherence, loras letting you play out any fantasy you want, and I2V minimax 3 shit that turns it into what looks like real video.As soon as we get local APIs that we can converse with and have them act as agents like Grok/GPT, it's going to be over for these companies. I can't tell you the amount of times Google AI and Grok have given me completely wrong answers, anyway, so fuck'em.It's only Indians keeping the online AI industry alive at this point, isn't it?
>>109956221>Why the fuck do people still gen on online models for?Because 95% people are retards. It's like asking why do people pay for Neflix or Disney+ when you can just download shows or films via torrent and you have them forever, and can watch them in higher quality whenever you want. You should expect a similar thing for AI. Even if the software was just a few clicks to install and easy to get going they still wouldn't do it because most are just cattle that only follow the herd and don't ever think independently or critically about anything.
>>109956221Give me a qrd on how to do itI have a 5070ti
>>109956545I'll give you the easiest way to do it;2 programs; Swarm UI is what you use to generate pics and videos, and OneTrainer is what you use to train your own loras of people when you want to start doing that.>Download "One Trainer"This program lets you make Loras of people you know (Loras are basically DLC for AI genning. Civitai.red is your friend to download Loras for things like clothing, poses, general concepts, etc). One Trainer has giant toggle profile switches to just automatically apply universal settings for what you're using; just apply the "Krea 2" profile and you're basically good to train on it until you get more indepth into it.Now download Swarm UI. This lets you gen stuff without jumping through hoops and using ComfyUI with its autism interface. You can just watch 10 minute Youtube vids / ask Google AI how to set things up inside it, and have it all sorted in an hour.Both these programs will automatically set themselves up and jump through all the aids loop of downloading dependencies and Python and PyTorch, etc. It's aids without it, trust me.For the 5070ti, download Krea 2 (lustify-v10-krea-turbo-int8_convrot) from Civitai.red. This thing will easily fit on 16gb vram, basically you ALWAYS want the versions that say "int8" and "turbo"; they fit on a 5070ti and let you gen insanely realistic stuff; each gen will take a few seconds, so long as you follow what Google ai / Youtube setup vids tell you what settings to do. For creating your own loras on OneTrainer, you'll need the raw Krea 2 model (just Google how to do it, the Google AI will walk you through it).If you can, bite the bullet and buy as much ram as you can (64GB is enough), go into Nvidia control panel, and under global settings enable "CUDA Memory Fallback" to "Driver Default". It'll stop you OOMing from the 16gb vram, but slow down your genning heaps. This will enable you to create loras for Krea 2 / MinMax.
>>109956640>>109956545The stuff I'm telling you is my own experience jumping through hoops doing all this shit, trust me, it's the easiest way. Do NOT get involved in "Comfy UI" or any other programs, they'll just confuse you and make you give up. OneTrainer and Swarm UI are retard proof and it's like using Steam but for genning and not managing your games.The reason you want to get as much ram as possible is you can offload your ram usage from Vram onto ram so you don't OOM (Out of Memory; the most common thing you can run into trying to local gen).It'll slow down whatever you're doing by heaps, but it makes it possible to gen basically anything. It's also required to create loras for Krea 2 and newer models on OneTrainer as they need the full RAW models to train on (they instantly use up 24+gb Vram, so the rest is offloaded into ram), but just leave it running overnight when you sleep and you'll have a lora of whoever you want to fap to in the morning.But honestly, any errors that get thrown up, just chuck them into Google AI; 99% of the time it'll give you a simple solution to fix it that otherwise takes hours of head bashing and listening to /g/ autist Indians.If you have more questions chuck them in here, and I'll answer them all in the morning to help you (Aus fag, 11:59pm here).So for now;>Download SwarmUI, set it up, go on Civitai.red, search for "Lustify Krea", download the Turbo int8 version, put it into the Models/Stable-Diffusion folder, then try genning, and whatever errors are thrown up, let Google AI fix it. Dependencies will all be downloaded automatically, dw. Also, Nvidia Control Panel, and make sure CUDA Memory Fallback is set it "Prefer no fallback" for now.Good luck for now, faggot.
