Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Stable Diffusion Public Release

stability.ai

Post reply on HN
Print view View original

141–150 of 437 posts

Re: Stable Diffusion Public Release

#141

Post by choppaface »

mabbo wrote: post #33

The most interesting part, to me, of a release like this is the amount of "please don't abuse this technology" pleading. No licence will ever stop people from doing things that the licence says they can't. There will always be someone who digs into the internals and makes a version that does not respect your hopes and dreams. It's going to be bad. As I see it, within a couple years this tech will be so widespread and…

Really, the licensing is most interesting? There’s a lot of public info about training and development too.

The license itself is pretty irrelevant. What people will actually do with the training blueprints, and how fast things will evolve.. now that’s interesting.

Re: Stable Diffusion Public Release

#142

Post by jsmith45 »

badsectoracula wrote: post #44

Is there any way to download this on my PC and run it offline? Something like a command-line tool like $ ./something "cow flying in space" > cow-in-space.png that runs with local-only data (i.e. no internet access, no DRM, no weird API keys, etc like pretty much every AI-related application i've seen recently) would be neat.

I think the answer is yes, but setup is a bit complicated. I would test this myself, but I don't have an NVIDIA card with at least 10GB of VRAM.

One time:

1. Have "conda" installed.

2. clone https://github.com/CompVis/stable-diffusion

3. `conda env create -f environment.yaml`

4. activate the Venv with `conda activate ldm`

5. Download weights from https://huggingface.co/CompVis/stable-diffusion-v-1-4-origin... (requires registration).

6. `mkdir -p models/ldm/stable-diffusion-v1/`

7. `ln -s models/ldm/stable-diffusion-v1/model.ckpt`. (you can download the other version of the model, like v1-1, v1-2, and v1-3 and symlink them instead if you prefer).

To run:

1. activate venv with `conda activate ldm` (unless still in a prompt running inside the venv).

2. `python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms`.

Also there is a safety filter in the code that will black out NSFW or otherwise expected to be offensive images (presumably also including things like swastikas, gore, etc). It is trivial to disable by editing the source if you want.

Re: Stable Diffusion Public Release

#143

Post by andybak »

jmfldn wrote: post #69
simonw wrote: post #4

If you want to try it out this seems to be the best tool for doing so: https://beta.dreamstudio.ai/

I just tried it via this link. I'm not sure what I'm looking at here but the results were extremely underwhelming. I've used Dall E 2 and Midjourney extensively so I know what they're capable of. Maybe I'm missing something?

Every model I've used initially seemed poor compared to the one I was just using. It takes time to figure out their sweet spot and what kind of prompts they excel at.

I've had a lot of great results from SD - but different great results to Dall-E.

Re: Stable Diffusion Public Release

#144

Post by eastbound »

gabereiser wrote: post #125
eastbound wrote: post #70

Earlier quoted context omitted.

But even today, we deal correctly with it. Fakes and real photos are mingled together in 9Gag/LatestNews reports about Ukraine. Under the fakes (and the real), people ask for confirmation. Someone says it’s true, no-one believes him, until a link to a newspaper is dropped. And 9Gag isn’t the highest IQ community around, so yes, general population does distrusts photos by default until proven. They are laughed at anyw…

The problem comes from the early adult newsroom interns responsible for sourcing content. They don’t know it’s fake, it sounds like a good click-baity article to them, so they run it. It happens.

I wouldn’t shift responsibility on the shoulders of the last newcomer. The top of the management has had ample time to diagnose this. If it remains like this, it’s by design.

Re: Stable Diffusion Public Release

#146

Post by mdcds »

mabbo wrote: post #33

The most interesting part, to me, of a release like this is the amount of "please don't abuse this technology" pleading. No licence will ever stop people from doing things that the licence says they can't. There will always be someone who digs into the internals and makes a version that does not respect your hopes and dreams. It's going to be bad. As I see it, within a couple years this tech will be so widespread and…

> you can fully expect your asshole friends to grab a dozen photos of you from Facebook and then make a hyperrealistic pornographic image of you with a gorilla

my prediction is that, as a result, people will start assuming pics online are fake until proven otherwise.

Re: Stable Diffusion Public Release

#147

Post by msoucy »

Sohcahtoa82 wrote: post #74
criddell wrote: post #54

Earlier quoted context omitted.

If I’m willing to buy a computer, any pointers on what I would have to buy? I’m asking for specific models from a company like Dell, Apple, or Lenovo?

You'll need a GPU. One with a LOT of RAM, like an RTX 3090, which has 24 GB.

According to this post, it needs 6.9 Gb. So the 3070, 3070-Ti, 3080, etc. can all run it. Sadly, my RTX 2060 is below that limit...

Re: Stable Diffusion Public Release

#148

Post by digitallyfree »

badsectoracula wrote: post #44

Is there any way to download this on my PC and run it offline? Something like a command-line tool like $ ./something "cow flying in space" > cow-in-space.png that runs with local-only data (i.e. no internet access, no DRM, no weird API keys, etc like pretty much every AI-related application i've seen recently) would be neat.

As an aside I wonder how performance would be like running this on CPU (with the current GPU shortage this might well be a worthwhile choice). Even something like 30 minutes to generate an image on a multicore CPU would greatly increase the number of people able to freely play with this model.

Re: Stable Diffusion Public Release

#149

Post by morsch »

aaronharnly wrote: post #109
morsch wrote: post #104

Earlier quoted context omitted.

In fact, they announced lower prices today, going into effect in September. https://openai.com/api/pricing/

That’s for GPT-3 text generation, not the DALL-E 2 image generator. Hopefully that will get pricing revised down (and an official API) before long.

Oh, I see! Thanks, I should have read the mail more closely.
Post reply on HN