- Joined
- Jul 25, 2019
Do you read? Its not crashing. Everything works fine but the previews and you can carry on without restarting the program.
Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump
- Thread starter Near
- Start date
-
🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
- Joined
- Jul 25, 2019
Anything past 2.0 removed celebrities I think. There was some other stuff about changing the dataset, but I don't remember. There are tons of independent and derivative checkpoints to use, so it doesn't really matter at this point. SD isn't as good as a more tailored checkpoint for most things.
- Joined
- Aug 7, 2022
trying for consistency in character design with different facial expressions, initial attempt went above and beyond my expectations
- Joined
- Jul 25, 2019
Either inpaint or using the same seed?
Edit: I seem to have produced a working hypernetwork. Not the best, but a good first go. Went faster than I expected.
Last edited:
- Joined
- Aug 7, 2022
Same seed, just different prompt. That's why the earring, clothing, and tattoos are different in all three.
- Joined
- Jan 10, 2020
How close are we to a net that can generate 1080p images on demand, maybe even up to 4k? In the same time window it takes current shit to make a 1000x1000 image, specifically. I know it's technically possible with some insane hardware, I mean for a few dollars a month or on a desktop gpu. Is the current resolution "cap" for consumer level image gen a hardware issue, a software limit, or a bit of both? I also don't mean upscaling to that res, rather raw output.
Edit: any suggestions as to an uncensored service that can dump higher than 1080p even if it's slow, that would work for me in the meantime.
Edit: any suggestions as to an uncensored service that can dump higher than 1080p even if it's slow, that would work for me in the meantime.
- Joined
- Dec 13, 2022
Right now the go-to is generating an image in like ~500x500 or 700x700p and then upscaling it, and given how the models are trained and gpu requirements you should just look for a two part process until the next gen of GPUs come out. Basically it's a consumer-level hardware issue AND a model issue, cause even top models are trained on <800p.
- Joined
- Jan 10, 2020
Why are the top models being trained on such relatively low resolutions? Would there actually be an issue trying to generate smaller images out of it at the same speed as one trained like that?
- Joined
- Dec 13, 2022
As far as I understand, while producing images takes a decent amount of VRAM, training models takes far more, and as you increase image size you're dealing with the inverse square law - i.e. doubling the image size means 4x the VRAM etc.
Not too sure I understand the question. Making smaller images is super easy. You could generate it at the regular size then scale it down, or inpaint in (or even straight-up generate) an image at desired scale.
The thing is you'll typically get best results generating an image at the same size the model was trained at - but with tools like upscaling and in(/out)painting you can still make compositions of basically any size with enough effort.
- Joined
- Jul 25, 2019
My user-level understanding is that it is a Vram limitation more than processing speed. Ask for too big of images and you run out. The released models are the ones correctly sized to run on home computers.
Edit: regarding training, I'm training right now and it's using 6gb of Vram on a 512 x 512 model. It crashed out immediately at a 1920 x 1080 attempt (only got 8 gigs). I'm certainly using more processing power when generating, or at least my PC is producing more heat and running fans more. Right now it's at 23% utilization. I'll check later when generating test images what it's pulling.
Edit: Edit: 37-42% while generating, still not really breaking a sweat (1080ti) vram maxed out, but the model for training is still loaded.
Edit:Edit:Edit: about the same vram to just produce images, so big shit would crash out in most home computers.
Last edited:
Girl Named Sandoz
kiwifarms.net
- Joined
- Aug 11, 2022
- Joined
- Apr 27, 2022
Last edited:
- Joined
- Dec 12, 2015
I was in a bit of a landscape mood.
- Joined
- Oct 3, 2018
Site breaks whenever I try to quote people , - but those independent checkpoints are all 1.5 based. LORAs are un-needed when working with 1.5 derivived models and only one generational application uses them (Automatic). It's kind of funny that you can get the same results if not better on NMKD without using LORAs if you use the prompts properly. I mean, it solves the 2.1 and onward problem - - but then it creates a new problem. People who rely on LORAS without actually learning how to weight their prompts correctly.
edit: I forgot to mention what they changed in 2.1
-Celebrities are removed and identifiable human faces are now harder to make. Inhuman people/ people with fucked faces/ non-traditional faces like Lena Dunham and Bella Ramsey still work. Fun Fact: 1.5 fixes Bella Ramsey and cures her Inmouths Syndrome. Loras for some reason retain it.
-Aging was made more difficult.
- Anything heavily copyrighted was removed, but traces are still there. It's harder to generate Chris Prat wearing a Mario costume now. Midjourney still does.
- More stock images were added.
- Pulls more watermarked images , stressing the need for negative prompts.
- Heavier emphasis on negative prompts and precision.
SD1.5 doesn't need precision to get winners, but needs a few more throws.
SD2.1 needs precision to hit the mark, but when it does, it's a winner almost every time.
Anyway
I've been trying to post this for DAYS, not sure if I should toss this in the seething thread as well.
Put a name to the artwork behind AI art with this algorithmically sma…
archived 7 Feb 2023 18:51:13 UTC
TL;DR;
Losers create "bot" that can detect Ai art.
Bot doesn't detect ai work. Literally, it cannot detect if an image was ai generated? What's the fucking point then?
Bot scans the image and does a "Similar too search" close to what google and tineye use and then it declares that almost every image inserted into it was
I inserted my dream diary image from a few pages back and wouldn't know?
Absolute ai forgery! Witchcraft! No, it's just retarded.
Last edited:
- Joined
- Apr 27, 2022
Lol. I played with that retarded shit for a while. It's definitely going to be the basis for frivolous lawsuits in the near future.
>mfw "artists" continue eating each other alive
- Joined
- Dec 15, 2022
One thing I've just now realized.
AI was put into man's reach because of gaming and crypto. Innovation in GPU's was necessary to get the critical mass processing to make AI systems viable.
AI was put into man's reach because of gaming and crypto. Innovation in GPU's was necessary to get the critical mass processing to make AI systems viable.
vladimirovanova
kiwifarms.net
- Joined
- Mar 11, 2019
The funny thing is that I still think art made by people is more interesting than art made by the algorithm. AI art is interesting because it's AI art and it's cool seeing what it can do and how useful it can be to have it on command if I want it. The knowledge that it's AI generated makes it more disposable to me. All these artcels freaking out is schizo because I'm never going to stop appreciating and consooming their shit.
- Joined
- Aug 28, 2019
And absurd, largely unnecessary amounts of VRAM in consumer cards (24 GB, probably 32 GB soon) directly benefits the off-the-shelf AI user.
- Joined
- Jun 23, 2016
It's the same idea as if someone gifts me a painting they painted themselves, you bet your ass I'm going to value that painting orders of magnitude more than if they had sent me a photograph of the same thing.