Okay, I did delete the last journal as having it at the tippy top of my userpage was giving off some too-heavy vibes; suffice it to say, please keep stuff you write on/about my pieces in the realm of fantasy and we will get along great. Sound great? Great! Thanks for that.
So I am like 10 days into fuckingabout with SD, the mini-journey has been interesting. Ostensibly one gets into it because they see a potential tool for unleashing or supercharging creativity in a way that trad art inherently limits (this isn't necessarily a bad thing btw - an argument can be made that art SHOULD be hard and laborious as it means you only pursue the ideas that really matter to you - but I digress). There is an undeniable creative thrill when you type in your first prompt and watch an entire art plop out the other end like a tiny miracle. This repeats when you copypaste in your first wad of negative and positive metaprompts (eg positive:masterpiece, beautiful, negative:ugly, low res, etc), when you plug in your first lora, etc.
Soon enough the limitations start rearing their head though. SD is amazing at creating lavishly photorealistic portrait of solo humanoid characters in beautiful locations who are posing uncomplicatedly, smiling and staring at the camera. As you deviate from that baseline, you watch the engine start struggling harder and harder to give you what you want. At a certain level of deviation it starts ignoring stuff it doesn't know much about entirely, and at a further certain point it just says fuck you and starts stuffing all your tokens into a nightmare katamari of limbs and abject suffering that doesn't really look anything like what you asked for at all.
Obviously this is what loras and similar plugins are for, to extend the model's knowledgebase so it can interpret what arcanine is or what a mating press looks like. At this point though you are kind of slamming all the way to the other side of the equation and replacing the model's vague idea of what you want with another human's (who built the training set for the lora) very specific idea of what you want; plug in your 'mating press' lora and now any prompt to do with mating presses will have the exact same pose, camera angle, expressions, etc just with different characters painted on top. It may just be my inexperience so far and I need to play around more with lora weights, multi-stage sampling, or whatever, but it feels like there's a creative sweet spot missing right now in the workflow, and it's the sweet spot I was hoping to be able to explore when I fired up SD in the first place. I really don't care too much about photorealistic solo portraits of X fan character with Y genitals in Z fetish situation, which appears to be what much of the SD community immediately gravitates towards and reproduces infinitely. I also don't care too much about utterly random and decontextualized dogcatcowwolf chimera people in random situations having random (and usually limb-melting) sexytimes together in a generic brown room. Which, lmao, is of course what I am generating and sharing right now as I feel out the countours and limitations of the program.
Anyways, there are frustrations, but it remains fun for the time being. I would say that from what I've experienced thus far, that the trad artist's case about SD not being a legit replacement for 'real art' is still a strong one. To me 'real art' is something that conveys a highly developed, strongly contextualized idea between one person and another via a visual medium. And SD just ain't there yet, lol. Unless the valuable idea you're setting out to convey is 'legosi and judy hopps in a room fucking in the style of reverse_cowgirl_v2.safetensors', in which case you're gucci.
| Viewed: |
73 times |
| Added: |
8 months, 1 week ago
30 Aug 2023 21:27 CEST
|
|
|
