Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
  • Informative
Reactions: indomitable snowman and AmpleApricots
  • Agree
  • Like
Reactions: AmpleApricots and The Mass Shooter Ron Soye
The Mass Shooter Ron Soye said:
if the rumor is accurate
Click to expand...
Eh, I buy it. The GPU people do not care about power consumption or building efficient cards and drivers. They simply do not. That watching a movie on my regular GPU takes about 20-30 Watts more than on e.g. a similar ARM system with the same software which consumes only 10% *in total* of that what the normal PC consumes *more* is testament to that. The CPU people get all the pressure in this space while people somehow just accept the computer sitting at 400W rendering a static vidya title screen because the GPU is caught at full blast in a tight loop.

I think the case with failure here will mostly be people paying with a literal bag with a dollar sign on it for the card and then getting the cheapest chinesium, highest rated power supply the least amount of money can buy because the card already was so expensive and their current power supply just won't be cutting it. I'm sure these IRNGX ATX 1000W power supplies are Number 1 quality.
 
  • Like
Reactions: Overly Serious and The Mass Shooter Ron Soye
The Verge: Google’s new generative AI video model is now available (ghost)
Veo, Google’s latest generative AI video model, is now available for businesses to start incorporating into their content creation pipelines. After first being unveiled in May — three months after OpenAI demoed its competing Sora product — Veo has beaten it to market by launching in a private preview via Google’s Vertex AI platform.
Click to expand...
Google says Veo and Imagen 3 carry built-in safeguards to prevent them from generating harmful content or violating copyright protections — though we’ve found the latter wasn’t difficult to bypass. Everything produced by Veo and Imagen 3 is also embedded with DeepMind’s SynthID technology — a kind of invisible digital watermark that Google says can “decrease misinformation and misattribution concerns.” It’s a similar concept to Adobe’s Content Credentials system, which can be embedded into content produced by the creative software giant’s own image and video generative AI models.
Click to expand...
 
macrodegenerate said:
I mean is there a better reason?
Click to expand...
Fine-tuning your own models is the one thing I would get one of these huge cards for, since they have the necessary VRAM. The amount of work you put it through would have to outweigh the cost of just paying to use a GPU on something like vast.ai though.
 
I noticed Amuse was updated to V2.2.2 and wanted to check out Flux.1-schnell but is it uncensored? I want to know before I download a 34 gig file.

I'm not a pervert, goddammit. I just don't want any censorship of gore or violence or anything.

Is Dreamshaper v7 worth a goddamn? Any ControlNet models to recommend to get my feet wet?
 
Does anyone have any experience baking LoRAs for SDXL or Pony Diffusion? I'm using an NVIDIA GeForce RTX 3070 that's 16 GB (half is shared) and my training times take around 4 hours with xformers and cache latents on with a batch size of 2; currently using a dim/alpha of 16/8. Is this normal?
 
🎁 🎶 Simply~ having~ a wonderful Christmas time! 🎶 ☃️
It's time for a festive avatar, so here are the ones I generated, and then the one I cleaned up
1.jpg2.jpg3.jpg4.jpgKiwiChristmas.jpg
 
  • Like
  • Feels
  • Winner
Reactions: ToroidalBoat, bru, Vulcanusii and 5 others
BrunoMattei said:
I noticed Amuse was updated to V2.2.2 and wanted to check out Flux.1-schnell but is it uncensored? I want to know before I download a 34 gig file.

I'm not a pervert, goddammit. I just don't want any censorship of gore or violence or anything.
Click to expand...
Flux won’t easily produce porn or gore for you, no. There’s a billion porn models already, just use one of those.
 
  • Like
Reactions: woowie queen and BirdUp
Susanna said:
Flux won’t easily produce porn or gore for you, no. There’s a billion porn models already, just use one of those.
Click to expand...
I'm more about gore. Nudity is incidental. Anything that allows me to make something like... This, I'm fine with:

 
  • Like
Reactions: Machine made Horrors
BrunoMattei said:
Is Dreamshaper v7 worth a goddamn?
Click to expand...
I have quite a few of the Dreamshaper series of models and like what they've produced, though I generally don't try for gore or photorealism if that's what you're after so I can't speak much for that. If I'm rendering anything that isn't some anime aesthetic odds are Dreamshaper is going to be at or near the top of the list of models I'll pick.
 
  • Informative
Reactions: woowie queen and BrunoMattei
BrunoMattei said:
I noticed Amuse was updated to V2.2.2 and wanted to check out Flux.1-schnell but is it uncensored? I want to know before I download a 34 gig file.

I'm not a pervert, goddammit. I just don't want any censorship of gore or violence or anything.

Is Dreamshaper v7 worth a goddamn? Any ControlNet models to recommend to get my feet wet?
Click to expand...
Tangentially is there a workflow that allows you to use control nets with flux? I thought I had one (now lost) working but the results weren't great and everything I'm finding now I think is generating with flux and then doing a second pass with stable diffusiom for the control nets.

Is flux still kinda inflexible? last time I tried using it was a couple of months ago.
 
Baraadmirer said:
Does anyone have any experience baking LoRAs for SDXL or Pony Diffusion? I'm using an NVIDIA GeForce RTX 3070 that's 16 GB (half is shared) and my training times take around 4 hours with xformers and cache latents on with a batch size of 2; currently using a dim/alpha of 16/8. Is this normal?
Click to expand...
Are you using steps or epochs for limits and how many? You may be over training or it just may be normal with a large data set.
 
macrodegenerate said:
Are you using steps or epochs for limits and how many? You may be over training or it just may be normal with a large data set.
Click to expand...
For this one I'm training by epochs (4). I've got 36 images that are supposed to repeat 3 times each, so I assume that's supposed to be a total of 432 steps, but my CMD prompt has it marked as 60 steps in total. Right now I'm clocking in at an average of just under 5 minutes for 1 iteration. I hear some people have had some success by modifying their bucket settings, but advice has been varied.

Maybe I should just do 1 repeat in the future with more epochs?

EDIT: Checked what NVIDIA driver I'm using and I'm at 546.12. I've been hesitant to update because of some of the things I've heard about newer versions being worse for training.
 
Baraadmirer said:
For this one I'm training by epochs (4). I've got 36 images that are supposed to repeat 3 times each, so I assume that's supposed to be a total of 432 steps, but my CMD prompt has it marked as 60 steps in total. Right now I'm clocking in at an average of just under 5 minutes for 1 iteration. I hear some people have had some success by modifying their bucket settings, but advice has been varied.

Maybe I should just do 1 repeat in the future with more epochs?

EDIT: Checked what NVIDIA driver I'm using and I'm at 546.12. I've been hesitant to update because of some of the things I've heard about newer versions being worse for training.
Click to expand...
Are you using the Kohya GUI?
 
It looks like Sora is a huge shit show because the censorship went into overboard. You can't do anything that even vaguely implies nudity (like say, legs or a bathing suit). Keep in mind that Sora has a $200 per month subscription.
 
  • Horrifying
Reactions: ToroidalBoat
BrunoMattei said:
It looks like Sora is a huge shit show because the censorship went into overboard. You can't do anything that even vaguely implies nudity (like say, legs or a bathing suit). Keep in mind that Sora has a $200 per month subscription.
Click to expand...
Sora access is part of the $20 GPT+ plan. $200 just gives you 1080p and unlimited gens which is worth bitching about, but if you actually expected less censorship from fucking OAI I don't know what to tell you other than you will be disappointed every time they release anything at all.
What you should bitch about, is that even with the $200 plan, you still burn through your credits absurdly fast. Even with the unlimited slow queue you get, it's still absurdly priced. It's better just to wait for it to get cheaper, because even though it's the best on the market, I really struggle to see the worth in it. I'm not sure there's a whole lot of people who work creatively who would try and use it yet at that pricepoint monthly.
 
  • Thunk-Provoking
Reactions: BrunoMattei
You must log in or register to reply here.
Back