AnyLoRA

The NOT PRUNED models are for using without making LORAs.

@Lykon Thought you said this "Remember to use the pruned version when training (less vram and no baked vae)." :-)

@Lykon Gotcha. So https://civitai.com/models/23900?modelVersionId=28562 is the pruned AnyLORA version you recommend people use to train LORAs?

Newest version is baked with Blessed_VAE

Got it,thank you.

@hazukita Why is that? Is it better than FT MSE?

@Lykon Gotcha, thanks!

I doubt they would do that. For whatever reason people don't like giving out their training 'secrets'. I dunno if it's some power trip thing or they just don't want to decrease their potential followers.

Lykon could be different but who knows. I personally don't see a point in keeping it private. 90% of people here don't give a fuck about commercial stuff and are probably here to view sexy anime girls

You can open a Lora in the webui (or a harder to read notepad) with the info button and pretty much have everything they used except the actual image database.

@Lykon Ah I see, sorry if I was presumptuous. I've never heard of Lina's notebook though if you could share a link. Is it a collab like https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb#scrollTo=OglZzI_ujZq-

@Lykon Weird, can't find anything about it anywhere lmao. Honestly the HollowStrawberry has been solid for me so far. Just made a Belle using your AnyLoRA and it looks great

@turkey910 it's right in the first line of the Strawberry colab you linked - "This is based on the work of Kohya-ss and Linaqruf. Thank you!"

Well, maybe I am doing something wrong, but there is a stark difference on my end between character or style loras that have been trained on NAI and those that have been trained on anylora, while testing on different checkpoints. I am not saying you can't generate great images with this model and I'd even say it's fine to test epochs that have been trained with NAI, but I don't see how you future proof yourself by training on a model mix, since you're going to lose compatibility. The claim about training from the description doesn't seem substantiated and I hope we don't end up with every lora being trained on a different checkpoint, which you're going to need to make effective use of the lora.

I will give them a try. Just on a side note, the biggest issues I had were not models made by you. There were probably some of them which had bad training parameters, but I doubt all of them had.

@Lykon I've actually found this to work fine for training people and using that to generate realistic or illustrated images of that person with a variety of different models

@Lykon 能说得详细一点吗，前面四个版本不知道如何选择

1D，2D，3D，4K畫質的清楚區別而已，不是誰都能使用最新版本，要看你的顯示卡，2060以下的使用2D，3060的顯示卡使用3D版本，4090才能下載4K版本，不然你的stable-diffusion-webui沒辦法運作。有些4K版本4090的顯示都會沒辦法運作。歐美國家的電腦配置是全世界最超前，歐美國家的電腦配置什麼版本都能運作。歐美國家不是使用顯示卡，他們是使用畫圖機器設備，裡面4090顯示卡安裝7-10張。稱為顯示卡合體的超級運算機。有些版本是給專業的工作者在使用的，比如廣告廠商，畫圖工作室，等等，他們不是使用家用電腦，是畫圖運算機，我們家庭用的，3D版本就夠用了，3D已經相當清楚，家庭用的，不需要下載4K。

also which one should I download, baked blessed or baked ftmse?

It does not depend on the model but on the content. It doesn't matter if an image is generated with a brush, photoshop or AI, if you play any content it's already copyrighted there's no way it belongs to you.

Internationally: An artistic work can only be protected by intellectual property if it has been created by a human being. The boundary of where you created it and how much was just the tool is blurry.

The use of AI in creative processes would not affect the protection of the resulting work under the copyright umbrella if, and only if, said AI has been used as a mere tool. That is to say, only in those cases in which the margin of human intervention is such, that there is no doubt that there is a natural person behind directing the final result, will the creation generate intellectual property rights.

@eypacha thanks for the info dude

The problem is the subject, not the image: if the subject or character has a copyright you can't use it for commercial use.

@markino That is dead wrong. Search Pixiv for images of characters generated by AI. People sell those images and no one does anything to them.

Pretty sure copyright is applied when there is enough human being has enough involvement into the art. So basically if you can do painting to change the image enough it can be copyrighted by you to an extent.

In my opinion, the random nature generated using the model tool creates new material from the model author's material. If the material can be artificially created twice, or polished by the team. It must be significantly improved in efficiency, but the premise is that it needs a soulful man-made design and an organizer to guide the creation. I hope this is the result, not a random flood of products. . .

@Saosaki and? there are people who sell copyrighted images, all the time people sell things that you "can't legally", how many fake brand products are everywhere, how many t-shirts with cartoons, movies, videogames, etc. copyright. if you take @markino estatement in a hyperliteral way, it is of course incorrect: you "can" even kill people.

@VitaminS randomness does not exist in deterministic systems. Algorithms are functions, all functions accept inputs and return an output. Same input, same output (always) Computers use pseudorandom number generator (PRNG) algorithms to generate initial image noise. The proof of them is simple. All images are reproducible. If you use the same checkpoint, same parameters, the same prompt and the same seed you will always get exactly the same image. People who sell images not posting the entries does not add randomness to the results

@eypacha This isn't entirely true. There's an ongoing issue that shows up all the time on the Automatic1111 and pytorch github issues pages where using xformers (or maybe it's fast attention? I'm on an RDNA3 card so I don't have access to any of that stuff yet) produces different images for the same parameters on different NVidia GPUs because of order of completion differences with varying CUDA core counts + greedy computation of results causing rounding errors. With the resources consumer GPUs have vs. the size of some of the matmuls happening you'd run into a big speed hit ordering computation of matrix tiles the same way every time. I got the feeling nobody really cares since we're talking about generating weird images from models, not calculating the trajectory and fuel use parameters for a manned trip to Mars. I certainly don't.

Then there's the issue of FP16 eventually degrading when low vram modes are enabled since copying FP16 from CPU->GPU and back loses the "invisible" precision.  The Comfy docs mention to restart the server every few hundred generations or so if you're running in FP16 and lowvram modes or results will start to become broken.

Actually I think I've seen the option to use random.org 's true random number generator based on atmospheric noise to seed the latents, or if you don't feel like connecting to the internet it would be easy enough to write something to pull numbers from the RDSEED instruction on x86-64 which pulls numbers off a non-deterministic hardware random bit generator that spits out 3Gb/s of garbage created from sampling silicon thermals and hashes it just to be safe. There's no repeating sequences there unless they're... well... random.

Then there's the thing where GPU and CPU "fast" versions of transcendental functions can vary... not to mention what backend you're using. I don't think these show up much in AI models except the schedulers but I'm not looking at any more python than I have to to get something working to find out for myself. I can't for the life of me get equivalent results in Comfy and Shark for the same prompt, seed, and image. Comfy uses DirectML on AMD with CPU offload for a couple of samplers (aka the ones to avoid), whereas shark uses a heavily optized llvm-mlir lowering that runs SD2.1 models faster on RDNA3 than any NVidia hardware can run 1.5 models (including the A100, I haven't seen tests from H100 yet) at the same resolution... assuming you only want to use a single LoRA and are willing to restrict yourself to the 3-4 sizes with tuning files, and are willing to wait 5 minutes to compile a new set of files if you change anything. I decided that until MIOpen is finished so PyTorch can start using RoCM on Windows I'll put up with complicated workflows I can do really cool stuff with in ComfyUI that take 60s per image vs the ability to spit out images every 1.2s but with very little support for changing much of anything.

Back to the actual topic the US Copyright beaureau already refused to issue copyright for AI generated images on that one comic book project and I'm pretty sure they already outright refused to issue copyright for it in the UK or EU. I'm sure somebody who generates with a model then extensively touches it up in photoshop would be covered most of the time, but I've had a couple of models spit out huge sequences of images for the same prompt that were clearly taken by the same photographer at the same location from multiple angles of the main subject (a seal). I told it to generate a "seal-cat with tabby fur" so it was effectively slapping fur and a deformed cat face on the photography. Unlike some cases where the model spits out random copyright / watermark nonense, these were all emitting the same watermark warped in exactly the same way. In some countries doing that by hand would be enough to be considered a derivative work but given how easily I spotted it I wouldn't have attempted to sell those without rights from the original photographer, whose identity can't be determined easily.

That or the series of cute cat girls in skin-tight bathing suits you generated with a realistic anime model ended up with a direct copy of somebody's 7 year old daughter's face on it and their dad finds it, pulls your address from the database of everybody in the known universe + address + emails that's floating around thanks to that MOVEIt breach and loads up his .338LM and hops in the pickup truck to drive cross country and blow your head off while parked 8 blocks away. For most people running stuff like this between gaming these days the obnoxious pulsating RGB lights in their computers will be enough to silhouette them against closed blinds so he'll have an easy shot as long as he only drinks one fifth of bourbon on the road trip instead of his usual two. :P

That's the real problem with it. Without reverse image searching the entire archive of everything that trained the model any given generated image could just be the original unmodified or barely modified image(s) depending on your prompt and how it was interpreted along with your other settings and you won't know until you get slapped with a treble damages suit on the going rights managed rate from a photographer or artist halfway across the country for selling "barely changed versions of their work with the watermark badly erased".

Until somebody trains a model from scratch that only uses creative commons / commercial use allowed and purchased royalty free images commercial use / copyright of model results will be a problem. That's why they don't provide the training image set for download anywhere, just the big list of URLs for you to fetch yourself. They're not allowed to distribute half of it. I'd be willing to bet, having been online for over 30 years now, that a run of forensics software over the drives they stored the training images on before sorting them would get Chris Hansen showing up asking them to have a seat. The SD models got away with incorporating as much copyrighted and probably some outright illegal imagery that the auto-filtering missed as they wanted to (they could have at least honored images with copyright set in the EXIF, y'know) because they were for "research purposes"... that whole thing falls apart when you start trying to do commercial things with it or anything derived from it.

Do a good painting based off your AI generated art with real acrylics and maybe I'd buy something from you (and we should really stop calling these gigantic state machine frozen networks "AI" until they're capable of continuously training themselves based on feedback and or extra image inputs as images are generated, and even then there's nothing like intelligence or thought happening at this scale).

That's just my take. Just have fun with it. You don't have to make money off of everything. It's like the damn youtube "content creator" thing. There was far more watchable content on youtube when people were just making funny stuff crap for fun and outnumbered the few corporate-sponsored channels. Now it's an army of identical used car salesman lookalikes who desperately need voice lessons giving terrible advice to try to make pennies off their video hits and an amazon referral for some cheap crap they're recommending. It's gone further downhill than The History Channel and Nick Nolte combined.

@GnomeExplorer I understand your point and I agree with a lot of what you say. I hope you agree with that classical computation operates within deterministic systems. When I refer to functions, inputs, and outputs, I mean fundamental mathematical concepts, beyond programming languages. The differences between different GPUs and other factors you mentioned are variables in this system.

Yes, I'm familiar with random.org and how they use sources of atmospheric noise to achieve a more genuine degree of randomness. This could be considered "truly random." However, the Newtonian premise would still hold that if we could replicate the same initial conditions, we would get the same mechanics and thus the same results.

Persi Diaconis, a mathematician known for his contributions to the study of randomness, has shown that coin tossing does not strictly follow the laws of mechanics when the physics of the toss is parameterized, and the same results can be reproduced in each toss.

On the other hand, Richard Feynman would argue that nature has quantum aspects. Following this line, there are programs like ANU QRNG that generate random numbers by measuring quantum fluctuations of the vacuum. This approach also provides an API for utilization.

I see that you're interested in this topic. Here's a link to a compilation I put together with material on the subject: I'm sharing this not to argue, but because if there's one topic I study and am passionate about, it's randomness. bit(.)ly /randompapers (civitai blocks the megafolder)

Is there randomness or not, it's a philosophically beautiful debate. As Robert R. Coveyou would say..."The generation of random numbers is too important to be left to chance."

Thank you for your contributions.