- This topic has 11 replies, 5 voices, and was last updated 1 week, 1 day ago by
.
-
Posts
-
I like how it enhances my base art but dislike how it decides things for me.
There is a point to where I feel it's not mine it's a kit bash of someone else's work. I'm already basically using models and environments someone else designed. Taking it one more step to just pushing a button and saying "make art" removes it even further away.
Creating a paragraph of what I want from the AI is no different than commissioning a piece from an artist.
You're not going to sign your name to a commission.I like the results for the most part. I just find it difficult to identify with the outcome and consider it something "I" did vs something "AI" did.
I think once we get to a point where we can have more control and consistency I'll warm up to it.
I know this probably sounds weird but I think there is something different about creation and replication. Maybe theres a fine line between being inspired and duplicating what youve seen but theres still a line.
I believe people process data a bit differently. They "feel" a color is right. They select a pose or a hair style based on what is attractive to them. I'd rather not have yet another product do the thinking for me.
I agree It's not art work to me ,using Daz isn't as well though IMO I just think it's kind of neat where it takes a couple of seconds to get an image made that some what has the look you were going for it has a long way to go before it will be really good because it's just in it's infancy right now.
There is always that grey area, it just shifts according to the person's perspective.
Think back to all the times there have been controversies in art history about art forgers?
IMO what they were doing took a lot more time & commitment then, than that spent by AI.
Arguably there was also a lot more artistic merit & talent, even if it was at best a copy.Creating a paragraph of what I want from the AI is no different than commissioning a piece from an artist.
Another way to look at it (or rationalize it) is: Does the term 'artist' apply more to the one who sits at the easel or the one who provides the 'vision'? The one at the easel knows how to paint people, trees, water, mountains, etc., that's the 'nuts and bolts' of artistry, the 'Craft'. On the other hand, the visionary, if they can communicate their ideas/emotions/'vision' to the artist, provides the impetus for him to interpret and portray those ideas on canvas.
So in that example, it took two to produce an original artwork. In the case of AI, there are some parallels to the visionary/artist paradigm, but it falls short in one respect: the visionary is human, as is the artist. They share the ability to visualize an idea and interpret emotions. The AI, however, is a contrivance that's limited to the model currently used. AI is not actually 'intelligent', but is generative to the degree that inference, sampling, training and model selection permit.
Still, to utilize AI requires appropriate model selection, lora picks, prompts that the model is trained on and the experience to determine how closely to follow prompting, like denoising and CFG. After those preliminaries, deciding on steps and batch sizes, knowing the resolutions the model accepts and iterating through several batches to finally pick out the 'keepers'...or...revamping the prompts and negative to more closely arrive at an acceptable set.
I wouldn't call that 'Art', but I would allow that it's 'Craft'.
I love that analogy @jorgen74.
Though this is not exactly pertaining to the issue, I notice the new NVIDIA 5000 series uses AI DLSS4 multi-frame generation to achieve high frame rates in games. Up to 15 in every 16 pixels rendered are not from the game itself but AI generated, ie. not real and invented by an AI algorithm. This way a frame rate of 25 FPS (without DLSS/MFG) can be pushed to 200+ FPS. This leads to a very dodgy claim by NVIDIA that the 5070 (12Gb) has the "same performance" as the 4090 (24Gb). A claim that has so many caveats as to be meaningless. I don't game so I'm interested to see benchmarks for Blender, DS, Flux etc. but I can't see the 5070 having anywhere near the "same performance" as the 4090 outside gaming.I agree it takes a sort of skill to direct the AI in the way you want. You select a general idea where you're headed just like someone selects an environment in Daz.
The weak link in the chain IMHO is the AI selects it for you. I ran some "art" I posted recently through AI. It selected the models face. Now I know there are controls to limit this but it still changed it. "It" is making changes I didn't ask for. It decides what it thinks I want vs me deciding what I think I want. In some cases it changes the clothing texture, skin tones, lighting ect... once the AI is done it isn't mine anymore.
@Frank: DLSS generates additional frames and is only effective for games that support it, just Games. AI needs VRAM and cores, Tensor-cores. DLSS is meaningless for AI.
Right now the 4090 is king for consumer-grade diffusion, the 5090 will have more of what the 4090 has, like 32GB of VRAM, more cores and more Tensor-cores. The price point going from $1.6k for 4090 to $2k for 5090 is more reasonable than I expected, considering the upgrades. But the street-price might be higher than MSRP at first, considering the usual scalper buy-ups at product launch.
And yes, the 5070 is basically a mid-grade card. It doesn't have anything like the performance of a 4090. In fact, the only thing it has is the ability to get fake 'vanity' fps up on the screen, but that doesn't equate to performance. What it *will* end up doing is imposing a 'delay' (latency) between 'real' frames which will affect competitive game-play in games that support DLSS4. A lot of gamers on reddit are not at all down with DLSS4.
@ADAM: That's kinda' what I was getting at. With AI, 'change' is sort of the object. I guess you could call it 'discovery'. You're sort of 'mining' the model to see what you can get out of it. Replication per se isn't what it's for, but you can limit what it changes in img2img by masking, like with inpaint. And, hey, if you're into it, you can try your hand at training your own lora to get closer to what you want, then you can merge the lora into the model you trained it on for your own custom model.Frank is probably the best guy to talk to about all this, he's got a lot of AI experience under his belt and could probably come up with a lot of tips we could both use, right Frank?
Thanks jorgen. I don't make loras anymore because I use Flux now and my hardware isn't up to it.
And with Flux, character loras don't work very well anyway. But of course Flux works better overall so I'm not going back to 1.5.
I've encountered a problem today where using PNGInfo and reusing the same seed does not reproduce the same image as before. It's really annoying.Might not be the seed, is anything different in the prompt? The neg prompt? The settings?Maybe pnginfo changed something or left something out in the extraction?
I wrote a png extractor in C to avoid any conflict or dependency on webui or it's python underpinnings, it's been working for me over a year now. You're welcome to play with it if you want. No UI, just right click on any png in explorer and it pops-up the metadata in Notepad++. Probably not as convenient as a webui extension, but I can use it regardless of whether I'm on A1111, Fooocus, Forge, or comfy, and I use all of them.
- You must be logged in to reply to this topic.
