Raw LLM Responses

As an AI user i got this video recommended to me, so I got curious and watched it. I train models, scrape data, etc. BUT I am also an artist; albeit I don't draw, I write. So, after a good watch, here are some actual tips: 1. Sign your work. That's it! Really! It's called metadata and you can add it on photoshop, SAI, Krita, etc when saving your file! (Also a visible one on your image). What? Expected more? Copycats have existed for centuries, people trying to paint a Mona Lisa and selling it as an original; artists imitating other artists to feed of their fame, art itself is the skill of replication. All pieces are replicated, all books have been told. (Don't tell anyone about Kimba and Lion King). What gives us value as an artist is our name. It's not the same a painting by Jhon Smith Smith that draws circles all day than one from the one and only Dali; and having a piece from Dali is by proxi more valuable than a copy from god knows who and looks "sortof" like a Dali. Art is and forever will be human, that's how it becomes art. (And why I personally refer to AI content as "gens" or "AI image" and not "AI Art". "B-But Name1 game uses AI art! That's a job taken from an artist!", Remeber, we as artists decide to turn our passion into a business, and by defacto you understand that businesses thrive on making more while spending less. BUT, what you also forget is that any business thrive on consumption, if no one buys your product you go broke. Then what? Give yourself a name so when I see the new Pokemon TCG deck done by AI or the new Magic expansion done by NameOfArtistHere, I rather have the magic expansion. Give yourself more value than the machine. Make ME want your content instead of the one made by a robot. That's what at leasts keeps me doing content, knowing that I can always do better. As for actual AI training, checkpoints are trained with huge data banks, the damage done by filters is null, Loras (small models) have been trained with purely toxic data to test the capabilities of the poison and turns out it just learns to replicate the filter as well, and the higher the "poison" the worse your image looks, so it works to train embeddings (filtering models) of what we don't want our generation to have. In other words your data serves as training of what we DONT want to get on our result. Here's how I train models: 1. Gather images, poisoned or not 2. Add regularization images (this is sortof a filter during training), if I got poisoned data on the 1st step AND as part of the regularization data this tells the AI not to learn the filter done by Shade/Glaze as it sees that said filter is not part of the character/concept/style 3. Train 4. Generate 5. Replace data with now filtered AI gens 6. Train again 7. Have fun The only real problem that is yet to be solved is removing artists' signatures. If you sign all your work then I dont have images without your signature as counterwieght (this is why I do step 4-7), and therefore the AI learns your name as part of your style, even more so, non-signed content then can easily be flagged as a replica and lose all value. How quirky such a small detail has protected us for so long. So if you are out there and are afraid, don't be. As long as you keep getting better, growing your name and yourself and creating, you will be ok. And if you want to make a buck then remember, we are selling ourselves; give the costumer a reason to spend their money in you rather than in a machine.