I saw a post about tracing up, and was curious what you guys felt about this?I've been using it a bit recently to make character sheets, or putting poses into it, and telling it to show the pose from a different angle, and it usually looks correct. I never trace, cause I actually want to get better at drawing, but I've basically been using it as a pinterest +.Whenever I see a pic from pinterest or elsewhere, but maybe the models im referencing only has a couple of pics, I'll just ask the ai to make a character sheet, or do a head sheet, and I'll get something that looks right. I dont use it a lot, but when I can't figure out how something looks, it usually points me in the right direction
>>7950645Is this really an AI reference? how about you do another character like Ashley or Grace?
>>7950645AI works in some cases and not others. They're good for creating references whose contents are well-represented in the AI's training set: muscles, clothing folds, fur patterns, landscapes, etc. They're also good at taking a reference and creating other references from it, like in your picture where the AI referenced existing material from the game. However, they're terrible at inventing new things since that requires understanding very abstract concepts and how they relate to each other. LLMs can kind of do this well now that they have "thinking" capabilities, but they're for generating text. Image generation models are a lot dumber since they can't think at all. AI likely won't be able to create appealing designs until image generation and text generation / thinking models are merged.
People who use AI as a reference that don't have the knowledge or skill to correct its hallucinations will inevitably be called out.
>>7950660"Thinking models" are literally just LLMs that are trained to explain a train of thought rather than a direct answer. So when it says it's "thinking" it's literally just outputting more tokens, that are hidden from you, and the rationale is the additional context will improve responses. They do have an equivalent to this for image generation which is essentially regional prompting where the model "rationalizes" where things go and the image is regenerated in certain areas. That's why Google's models have such good prompt adherence. What you're describing, or what I think you're wanting, is not possible with LLMs or image gen. Every model is constrained by its training until they come up with a way to alter weights based on continuous input and that may not be possible with the hardware we have or will have for a very long time.
>>7950662Yeah I tried to use it for faces and eventually I realized AI always fucks up the eye and ear placement. The rendering makes up for it
>>7950645Messed around with it, it:s still not good enough. By the time you are good enough to notice all the mistakes it makes, you won:t even need it for references. You can tell when certain artists are doing that, their styles all become more average and like AI. Not sure why you would want your art to look like AI.
>>7950645I don't see a problem with it but fix the mistakes it makes.
>>7950645Why would you do that when real life is more beautiful?
>>7950645It's a bad idea that's been elaborated on a dozen times by now, nearly every time this thread is posted. So I'll tell you how I feel about people asking - fuck off.
>>7950660Because AI is what you train it with is what you getIf you have images of 3D renders in the dataset, if you prompt for those items, you'll get those output.And as other anon says, "thinking" model isn't really the model thinking, it's just another LLM that takes your input and then elaborate on what your input should be before inputing it into the main model. Think of like, if your input is "write a 200 words story about a furry wrestler", instead of just taking your input as is, it'll elaborate/detail out on what the input should be like adding the character details, the setting, the plot points and so on before parsing it to the main LLM. They do it this way because simply making the models bigger don't improve on the performance of these models anymore if the input prompts are just shit single-liner queries with zero details. But the more detailed and concise your prompt is, the more the LLM would have to make connections to the output that will likely match to your query.Also all these image models from OpenAI, Gemini, Grok are more or less just a passable photoshop alternative that you just write text for. It's not great for art stuff, but I used it a bunch of time for some sales presentation where if I'd do the diagram myself (like a flowchart or conceptual machine layout) it'll take like 2-3 days for a presales shit that's not guaranteed to get a buyin. But as far as using them as refs, it's only useful if you already know how to cross reference real refs. It's shit if you don't know what's right or wrong.
yall are definitely right that it makes alot of mistakes, but yeah, if ya jsut use it abit, I havent; found any problems with it. >>7950660is right. I dont wanna sound like ai shills, but it sorta can be used as a tool , if you know how to use it.>>7950653heres picrel
I also just asked it to create some other outfits for her, with similar themes from the 2000s, and honestly I think their pretty decent.Parts of them dont really make sense, but I feel that just being able to mix and match things, even just to set a vibe, can be powerful. I mean honestly I love the clothes in the second one, the colors are actually coherent.Alot of people just get clothes from pinterest, whats the difference in also using ai to find stuff?Im asking seriously btw, Im actually curious
>>7950927What model and lora is this?
>>7950941literally just chatgpt+. I got the premium model free for a month
>>7950945Chatgpt had me and my coworkers using it for some simple sales pitch diagram/illustrations which it can do convincing enough if you add "illustration purposes only". Where you just want to show or explain to potential customers the overall idea, but not get into details yet until an order is given. Works pretty well for that part. Wish this model is open sourced and able to run locally though. 4 images a day is basically like welp that's it.
If AI is bad, then why is it good?
>>7950963Funny that, swine that eat from the trough ask themselves the same thing
>>7950974And I eat the swine. What is your point?
>>7950963Satanic indian/jew shit, don't care how good it is
>>7950963Why did her proportions become that of an 11 year old's in the bottom right? Matter of fact, wtf is going on with that tail? One angle lmfaoo
>>7950963simply because it copies and can't create anything new.
>>7950645It's not necessarily bad, but there's something about AI that often looks off, and has this very... well, 'AI' feeling to it. I think the last thing you want, is for your work to look like AI, not because it's so good, but because it has that certain "off feeling" AI tends to have.Also, if you're one of those people who a staunchly against AI for moral reasons, they may dislike you work for just using the reference, despite there being no actual AI in your work.
>>7950645I don't use AI, I will never use AI, and I will not associate myself with anyone using AI in their art making. There's no reason to use AI as reference.
>>7950963What's that Tolkien quote people love to throw around?>"Evil cannot create, it only corrupts..."Sounds pretty apropos for AI in the arts.
