>Finally decide just ignore 99% of the "advice" given on LORA training and just autistically created my own method over 4 months worth of weekends sitting infront of the trainer, with Youtube on in the corner, and re-trained multiple models from every epoch @ 150 steps until I perfected what I wanted at 6000 steps that had zero over-fitting and rigidity, despite every single "expert" screeching that you cannot train a LORA that high without it being a brick wall>Put models up on Civitai, become number one new creator>Everyone ranting how good my models are, have big creators messaging me asking what my method is and my settings>I'm too stupid to actually know what 99% of the stuff in training does or means, as I just adjust random numbers to make the results better, so I just read several settings from Onetrainer to make myself sound smart>Tfw Yogi just followed me and gave me a tonne of buzz because he liked my models>Have one of the top 3 ask if he can pass my training data off to the guy who de-censored Lora 2I finally found my calling in life. To haphazardly fail upwards with style, clothing, environment, and character LORAs for people to gen the most vile shit possible into. I don't really know how to feel about it.I legitimately feel like Rain Man at this point, only more retarded. Should I give my training logs and configs out to the big creators, /g/, or horde it? I want to keep creating models and people seem to worship them, but I don't want to become one of those profile retards who sit on that site all day begging for likes and buzz.Also, has anyone else been in this situation? I've tried going into the Local Gen / Diffusion generals on here and have no idea what most of the stuff people are talking about even means, then I see their work and it looks like absolute blurry trash.
Did you think your shit was worth a thread?
>>110107201Is it a loli?idc otherwise
Cunny or No Cunny?
Could have answered me OP
and yet it's still visibly and painfully slop for braindead subhumansdo AIniggers really?
>>110107201How much money do you make?
>>110107201>top LORA creator on Civitailet me guess loli breast expansion futa fart vore
>>110107201release it bro, you'll have a good name, otherwise the next "big thing" will take you out and then you won't have a good name
>>110107201Why does this matter?
>>110107201If you release it, it makes the overall quality in the community go up.
>>110107463>>110107479
almost like art is about being creative and doing your own thing rather than strictly "following best practices"
>>110107201Yes, I was in your position for audio loras and custom voice model finetunes for people doing live sts. I got bored and just do commission stuff now, I refuse to share publicly with ungrateful jeets and chinks anymore
>>110107201>on civitaigood morning sir
Civitai cannot even load a picture without needing a page refresh to close it again, why should I care.
>>110107959You still need to learn the classics before you can claim shit you make is anything remotely good art. Picasso knew how to sculpt for example
>>110110430not necessarily, art becomes good art if people like it
>>110110436Nah, you need to know the rules before you know what to break
>>110110456i'm not suggesting it wouldn't help. i didn't say to break the rules either, op didn't break them, he just went for it without really understanding either way.not usually the most efficient method, but it can work.
>>110107201nice bro. good shit
>>110110467What? OP stated that it took him months of learning
>>110110481he spent he fucked around in the trainer, but not strictly that he learned from someone else. while he did say he "had youtube on in the corner", he didn't say what kind of youtube. he wouldn't have taken that long or ended up with a strange result if he was just following guides
>>110110497Lol, everyone needs a touch off point.
>>110107201so how much money you get from all thisplease dont be a homosexual and tell me zero
>>110110504best way to find out what a button does is to just press it and see what happens. if you're just taught to not press it you might never press it and just miss out on what it does
>>110107959>prompting AI is doing artlmao
>>110107201so your method for making loras is messing with the vars randomly until you get what you like?
>>110110521Yeah lets push the nuclear button
>>110107201Somewhat, I'm reverse engineering some pretty influential software using qwen 27b (bonsai) on a 16gb 9070xt. Somehow I managed to set up a hermes workflow so efficient it flies past people trying to oneshot with the massive cloud models. All I've been doing is stopping it occasionally to review how it's doing and whipping it to improve every step. I have no clue what the fuck it's doing desu. It's mostly low level memory tinkering. I'm not making any money or a name for myself. I prefer anonymity instead of getting hunted down by lawyers from big companies.
>>110110551do not upset the machine spirits
>>110107201>Should I give my training logs and configs out to the big creators, /g/, or horde it?>should I work for free even cuck wagie cagies get paid for what they doit does feel good to have people praise you and use your stuff, I made some mods that got a lot of appreciation and I know it feels goodbut keep in mind civitai is another place controlled by kikes who make all these tools available for the community to do shit for free so big corpos can steal it and profit off it, just like linux, other open source applications and scientific research. if you are ok with that go ahead. I would also consider if you can use this in your favor. can you show this in a portifolio and linkedin? I am assuming this involves a lot of sick porn and, if so, I would hold up and produce sfw content with it and publish it as sfw content so you can use in your linkedin to get a job. if you release it now and can't link that hentai producing persona to your linkedin it will look like you just copied some cunny loving freak on civitai
>>110110567>pretty influential softwarei know that feeli am currently editing explorer.exe to show pics of big 3D titties instead of regular icons
>>110107201Whst caste are you OP
>>110110599>i am currently editing explorer.exe to show pics of big 3D titties instead of regular iconsI thought I contracted rotten goonerbrain but I think I'm good
>>110110613internet is for porn no matter how you falsely LARP as no porner
>>110110552You jest but "let's fuck around with live targets and find out what really happens long term" was one of the major reasons the nuclear button was pushed that one time.
>>110107201This is bullshit, but i believe it.
>>110110567>bonsaiYou have a $1000 GPU but not even 64GB RAM for QFN? 27B wastes too much time whipping itself on every step, while also being slower than cached A6B.>>110102090>>110102119
>>110110890That's no longer true though when you can gen locally offline
>>110111250I have 32 gb sys ram. It does seem very insecure compared to other models. For work like this I do prefer it's a bit insecure instead of going schizo when it pastes something wrong off memory into a file. One wrong digit in a hex fucks everything.No clue what QFN is, I've been focusing on the workflow itself mostly instead of optimizing the model. I'll check it out, thanks anon.
QFN is Qwen Flash Next. A big download with little memory footprint thanks to n-gram embedding structure.
>>110111338>>110111375Oh and the recommended engine for it is Strata, it has platform and model specific tweaks that llama refuses to implement because they do not apply to any other model.
>>110111375>>110111386hmmm. I want to use my puter while it's running though. The program it's decoding takes up some space too so I wouldn't have all 48 available for the iq3_s. I'm already running the bonsai 2 27b ptq1 with a 130k ctx window and 35 t/s output average comfortably. I don't even notice that it's running besides the fans and occasional coil whine. That's running on strata backend too. It's been running uninterrupted for 21 hours now making steady progress so I'll have to investigate how the two compares and whether it would pay off to switch models when I'm not at the computer. Thanks again!
>>110111464I was wrong, it's not strata compatible. If those numbers QFN numbers are true it would be a huge boost in speed, jeez.
>>110111555Bonsai is a dense model, meaning it has to activate all 27B weights for every token.QFN is mixture of experts, meaning it first selects what "expert" is relevant for each token, and then activates only 6B weights. At similar quantization that's 4-5 times smaller data movement per token.Furthermore, for any given prompt only a limited set of experts stays consistently relevant, for example "english language", "programming", "assembly" - these experts are cached in fast vram. If your task pivots into different skillset, like if you ask it to document what it's done, it can unload "assembly" and cache "markdown" instead - a one time slowdown that soon returns to top speed.
>>110107201You should give your training data. That's the most important. Your loras will become obsolete in a few months/year when a new base model will come out, but using your data someone can make a new one.I don't know how though.I'm also a lora maker and I obsessively collect pre-AI images to train my loras. I don't know where the fuck to upload it though because it's borderline porn and copyrighted.
>>110110890I am by no means a noporner but you are sick bro
>>110115223The biggest piece of advice I can give you regarding agentic work is to make your agent meticulously scan the internet for known solutions and skeletons. There's no worse feeling than spending hours/days on something, only to discover a chink already had already done it ages ago and posted it online for free. Always steal, always modify existing work, never reinvent the wheel.
>>110115323Thanks, I appreciate the tip
>>110115223I train specific subtle concepts like a type of makeup or a particular expression.Let's say you want to generate a type of ears.The gist of it:-find a keyword that generate random stuff in your model, like caliears8373. It should absolutely not generate ears.-create a GOOD dataset, preferrably non-ai images of the concept you want.-describe each image mentionning everything BUT the ears. Some models wants tags, some description, some no description at all. The newer wants long, good description. I caption with a local LLM telling it to NOT mention the thing.-on Civitai it's like 3$ a run, the number of steps depend on the models, use at least 768x768 pics. On runpod using "ai toolkit ostris ui official" as a pod template it's a bit cheaper and more reliable: rent a gpu (some newer gpu won't work so to be safe rent something a bit oldet), connect by the web interface, create a dataset, create a job using this dataset, choose the base model. Save the last 20 epochs.-the number of epoch depend on your model. I usually do 15 or 20 epoch even if it's overkill.-In both case you'll have to define exemple pictures that will be generated each epoch: "A woman with caliears8373 is sitting on a chair..."-run and you should see the images get more and more like your exemple.Then you have to test all the epochs, not just the last, because epoch 7 may be better than 6 or 8. Use different strenghts, 0.8 to 2.If there's something bothering you like a yellowish tint or some blur is probably because you have an image with this tint or blurry in your dataset. Remove it and retrain.Usually it takes between 5-10 attempts or correcting the dataset/the captions/the epochs to get a lora I'm happy with.The exact parameters are highly dependent on the model, you have to try.
>>110107201I have no idea how to set up a machine to run AI, but have faith one day I can just buy one, in the meantime I'm looking at my comics collection, so my question is: could a LORA workflow be able to recreate a style? Idc about copy right, since certain details can be changed, but is it possible?How would such a workflow even work? should I scan my comics, feed to LORA and it being able to recognise shapes and create a model based on that input? Or should I explain somehow what those shapes represent frame by frame?Any clue would be appreciated.
>>110107939What model was this made with?
>>110107201Yes, you should publish it because it will help the entire community be better.You will still have a job since like 1% actually do LORAs training and many people don't even have a GPU anyway.
>>110116076moss-image
>>110107201nobody knows anything, hence programming being bulldozed by AIeveryone in this world is a liar
>>110107201>>I'm too stupid to actually know what 99% of the stuff in training does or means,That's the same for everyone. LLMs are a black box and there's only people pretending to understand it as they throw stuff at the wall thousands of times.
>>110116064There is a lot of loras like this. Go browse the models on civitai.Ex: https://civitai.com/models/2202620/ed-benes-styleYou have your steps backward, you should first choose a base model (something easy to run, diverse, popular, easy to train), make a dataset (=collect the pictures of your style) and use civitai or ostris to create a lora (=a small file that'll connect to the base model to steer it toward what you want).This tutorial is a good start: https://www.promptingpixels.com/tutorial/z-image-turbo-lora-training
