128
If they charge by the image, and there is any expectation that AI image-generators might produce garbage results, then AI providers are incentivized to make you pay for their experimentation (lemmy.sdf.org)
submitted 6 months ago by xia@lemmy.sdf.org to c/showerthoughts@lemmy.world
16 comments fedilink hide all child comments

without you ever knowing about it, as well as (perhaps) swapping in a cheaper-to-operate model some percentage of the time, perhaps as request loads peak, hoping you'll just roll the dice and try again.

top 16 comments
sorted by: hot top controversial new old
[-] K3zi4@lemmy.world 27 points 6 months ago

I'm still unsure why anyone would pay for AI image generation purely because of the trial and error it takes. I get that not everyone has a GPU that can do it, but I use stable diffusion through automatic 1111 and I'll likely be about 2-300 generations of text to image, image to image, some inpainting and editing, then some more image to image and upscaling before I get a representation of what's in my head down.

I love the process of it all, but paying for tokens would completely limit me. Is there a specific reason that people use paid models? Or is it just because a lot of people are limited by their gpu?

[-] Doombot1@lemmy.one 11 points 6 months ago

GPU limiting and a general lack of either knowledge or wanting to put in the effort to do it themselves. Even just going into Github in the first place is enough of a barrier for a lot of people, unfortunately

[-] Iceblade02@lemmy.world 7 points 6 months ago

Yep, I've a mobile 3070 in my laptop, and whilst I feasibly could run some of the smallest models around, paying on a per-use basis gets me way better quality results for relatively cheaply.

Besides, running it locally isn't free either. Your hardware deprecates and depreciates over time, in addition to non-negligible power costs.

[-] chicken@lemmy.dbzer0.com 4 points 6 months ago

Local stable diffusion takes more effort to get good results, I think that is the main reason.

[-] Sakychu@lemmy.world 2 points 6 months ago

Sounds interesting! Do you maybe mind sharing some resources to get started with that?

[-] K3zi4@lemmy.world 2 points 6 months ago

Not sure if YouTube links are allowed here, but if you just search for automatic 1111 install tutorials, Sebastian Kamph's channel is excellent and pretty much everything you'll need to know is on there!

Keep an eye for his updated videos though, as you can imagine the tech moves so quickly that some of his more popular videos have been remade and updated!

[-] Sakychu@lemmy.world 1 points 6 months ago

Thanks, I will give it a look!

[-] Doombot1@lemmy.one 2 points 6 months ago

Easiest GitHub install I have ever done - works well on both Linux and windows per my own testing. Go onto the Automatic1111 stable-diffusion-webui github and find the git clone link, and then, from a command shell, use “cd” to get into your directory of choice and then do a “git clone ”. After that, go into said folder and literally just run the webui.bat/.sh file (might be called webui or might be called setup, can’t remember). That’s it - it should install all of the packages and python stuff and everything for you, and from there, honestly just fuck around with it for a while. On Linux, I had to install a few extra things/screw around a bit to get it working fully, but it wasn’t too tough. I highly recommend also installing the “dynamic prompts” extension and playing around with wildcards and such. Civitai has some great models and LoRAs you can play around with, too - Dreamshaper is a great one. If you’ve not got a 4000-series GPU, I would suggest editing the webui-user.bat/sh file and adding “—medvram” or “—lowvram” to help you get higher res images at the cost of a little bit of speed. Let me know if you’ve got any more questions! If nothing else, it’s honestly just a lot of fun to use.

[-] Sakychu@lemmy.world 2 points 6 months ago

I didn't expect such a detailed answer: Thank you very much! I will definitely check it out! And I will definitely come back here if I have questions!

[-] Doombot1@lemmy.one 2 points 6 months ago

Sure thing! I tend to over explain things anyways, lol

[-] where_am_i@sh.itjust.works 1 points 6 months ago

oh, look, it's one of those guys who grows his own potatoes too.

[-] K3zi4@lemmy.world 1 points 6 months ago

Hahaha, I hadn't considered it that way. Touché.

[-] Daxtron2@startrek.website 10 points 6 months ago

Part of why it's so important to have open source models

[-] kevincox@lemmy.ml 6 points 6 months ago

Yes, but in theory this is factored into the price. If you get 1/4 good images then a service that provided 4/4 good images could be 4x as expensive. If a service lowers their quality to push people to regenerate more often in theory those people will see their costs going up and quality go down and switch to a competitor.

This space is pretty competitive. So while tricks like this may be used in the short term to obscure the actual cost of using the service this will probably be noticed by reviewers and customers and will likely balance itself out over time.

[-] RagnarokOnline@programming.dev 5 points 6 months ago

I think there’s a chance we’ll see that type of cost-cutting eventually, but the only way I can imagine a pay-per-image model working is if you get unlimited “spins” until you get an image you find acceptable.

[-] lurch@sh.itjust.works 4 points 6 months ago

AI gambling

this post was submitted on 21 Apr 2024
128 points (93.8% liked)

Showerthoughts

29525 readers
774 users here now

A "Showerthought" is a simple term used to describe the thoughts that pop into your head while you're doing everyday things like taking a shower, driving, or just daydreaming. The best ones are thoughts that many people can relate to and they find something funny or interesting in regular stuff.

Rules

founded 1 year ago
MODERATORS
vatlark@lemmy.world
forkball@lemmy.world
quinten@lemmy.world
Thekingoflorda@lemmy.world