Dall-E AI

Want me to crash it?   

Quote

Neural nets need to be selectively trained to give the outcomes you want.

Also I should note that it is not the case that all neural networks need hands on training. This is a subset of the types of neural nets that require supervised training. Data selection is of course important still in those cases.

Yes, you can just run a neural net if you set the outcome parameters correctly from the start. Some setups can do quite a lot of amazing stuff with almost no input. But I dont think this is easy with something like this.

Quote from: Username on July 13, 2022, 04:52:06 PM

Quote

Neural nets need to be selectively trained to give the outcomes you want.

Also I should note that it is not the case that all neural networks need hands on training. This is a subset of the types of neural nets that require supervised training. Data selection is of course important still in those cases.

Yes, you can just run a neural net if you set the outcome parameters correctly from the start. Some setups can do quite a lot of amazing stuff with almost no input. But I dont think this is easy with something like this.

You could also automate your supervised user. There is already an image -> text service available. Since these are competitive algorithms, you could make your decision on which to choose based on which text translates more closely to the original input.

Also it doesn't really have to do with the 'outcome parameters' but more with what you are doing.

If you are looking for similar things, clustered data, things that are abstractly associated with each other you are likely using unsupervised.
For classification (is it a hotdog or a penis? is this picture an apple? etc) or if you are trying to understand relationships between different features or variables you are likely looking at supervised.

Also supervised does not mean a person needs to supervised. The data can simply be labeled. In the case of images like these used for training, that could be as simple as using its alt tag, throwing it into Rekognize (or whatever its called), or using the search term that brought it up.

This is probably supervised because its predicting data based on input, but I would guess its not supervised by a person but instead by something akin to a search engine.

Also it doesn't really have to do with the 'outcome parameters' but more with what you are doing.

If you are looking for similar things, clustered data, things that are abstractly associated with each other you are likely using unsupervised.
For classification (is it a hotdog or a penis? is this picture an apple? etc) or if you are trying to understand relationships between different features or variables you are likely looking at supervised.

Also supervised does not mean a person needs to supervised. The data can simply be labeled. In the case of images like these used for training, that could be as simple as using its alt tag, throwing it into Rekognize (or whatever its called), or using the search term that brought it up.

This is probably supervised because its predicting data based on input, but I would guess its not supervised by a person but instead by something akin to a search engine.

I have only dabbled in ML, and done mostly lots of reading on it. But my impressions in the same as what your saying.
If you want more specific outcomes, you do need more tuning, which is harder as the definitions become tougher.
A lot of the images we see are generated by pixel enhancing algorithms. Basically the first algorithm gives us a 64x64 pixel image of something that it relates to the input. Then the next round of work is just guessing how to fill those pixels with more pixels based on the context. The result is impressive, but its still pretty random.

Either way, I dont know if I should be scared or excited about what applications this software will be used for going forward

Yeah it certainly raises some questions. I think I am excited. I think what worries me more than malicious use (like I don't know, making folks naked or something) is misuse by folks that don't understand what it's doing, namely the police. World powers though certainly could leverage it, but I would think attempts to do so would at first be easily identifiable.

I took some ML in university and have since taken some classes on the mathematics involved as refreshers / to keep me up to date. I actually almost switched my career path about five years ago but decided I'd be happier / more stable in a corporate job compared to a start up. The math behind a lot of it is pretty nifty.

As far as applications I've played around with a few, implemented some as well as used tools like tensor flow or those provided as a service by various cloud providers. So I'm definitely no SME.

Yeah it certainly raises some questions. I think I am excited. I think what worries me more than malicious use (like I don't know, making folks naked or something) is misuse by folks that don't understand what it's doing, namely the police. World powers though certainly could leverage it, but I would think attempts to do so would at first be easily identifiable.

I took some ML in university and have since taken some classes on the mathematics involved as refreshers / to keep me up to date. I actually almost switched my career path about five years ago but decided I'd be happier / more stable in a corporate job compared to a start up. The math behind a lot of it is pretty nifty.

As far as applications I've played around with a few, implemented some as well as used tools like tensor flow or those provided as a service by various cloud providers. So I'm definitely no SME.

Bad things I see coming out of this.
- Fake evidence of anything you can think of. Political murders, simple footage of people steeling, countries bombing civilians. Countries not bombing civilians. Basically, we will end up trusting online news less than we already do.
- Lots of naked celebrities.

Good things
- you type a paragraph of they type of movie you want to watch, then get your own custom movie.

The one where flash floods cut off the two bottom floors and Rachel gets 28 days later rabies.

I just got my invite for the beta. Its pretty nice.

I just got my invite for the beta. Its pretty nice.

How long did you wait for the invite?
Its apparently quite a wait.

Quote from: Username on August 03, 2022, 10:11:48 AM

I just got my invite for the beta. Its pretty nice.

How long did you wait for the invite?
Its apparently quite a wait.

About a week or two. I didn't sign up until I saw someone else got into the beta on facebook a week or two ago. I believe its because I used my work email. I signed up with my personal at the same time and it has yet to be fulfilled there.

It seems almost impossible to get it to draw a flat earth lol.

It seems almost impossible to get it to draw a flat earth lol.

Further proof the world is round!

About a week or two. I didn't sign up until I saw someone else got into the beta on facebook a week or two ago. I believe its because I used my work email. I signed up with my personal at the same time and it has yet to be fulfilled there.

I never signed up because I thought it would be a full month wait. Seems I may have been right. But I should just sign up anyway.

So I signed up to Dall-E and im still waiting for it.
In the mean time I signed up to Midjourney, which is another version of it. Lots of fun, but not completely free. There is also Stable-Diffusion, which is a bit more effort for less perfect results.

So I looked for some images of flat earths that others did, so these are not mine. And will drop them down here for inspiration to you flatties. I will include the text that motivated Midjourney to create these.

flat earth, fermiment


flat earth, realistic


a horizontal Ancient Greek “flat-earth”

+
floats center scene as a table for Klothys goddess of destiny, she lords above it angrily as her wide symmetrical curly horns

+
hold her long white jumbled hair obscuring her eyes, swirling clouds of neon purples oranges and greens surround the “flat-earth” and fill the background,


the celestial sphere revolving around a flat earth


first flat earth map ever made


super realistic rendering of the flat earth depicted as if it was an iceberg with tree roots exposed beneath it


detailed flat earth from the Bible in a fishbowl


oil painting flat earth, ancient, modern, vibrant colors, vintage ultra realistic


a picture of the flat earth and dome, sat on the back of four turtle swimming through space, realsitic 8k


I found these to be the more interesting ones.
Hope you enjoy them too

Very interesting.  They don't have the burn victim aesthetic like Dalle mini does

Very interesting.  They don't have the burn victim aesthetic like Dalle mini does

Dalle mini is a old version of the technology. By old, I mean its about a year old.
The pace of development here is scaring the socks off of me.

In just the last month, they have halved the processing power it takes for Stable Diffusion (one version of this tech) to generate an image. In one year from now, these images I just showed will be laughed at. Midjourney, that created these images, is still to include newer methods developed in just this month. We are about 1 - 3 years away from full feature movies being generated directly off a script you feed it. Heck, just a few descriptions of what you want to watch.

This is going to be like reliving the rise of the internet. Fun times

Dall-E, not mini but the real thing, just removed their waitlist.

I'm trying to get it to do something interesting.  The quality looks better but it doesn't quite make sense out of what I'm giving it.  Also it looks like it intentionally blocks making art about some people.  It wouldn't let me generate this description "dracula dancing with alexandria ocasio-cortez in an Applebee's".

Yeah. Dall-E has a lot more limitations than Midjourney. Stable diffusion has even less.

The OpenAI crowd are worried about people using it as a weapon or to mess people around.
They are trying to be morally correct and all that. But the simple fact that they do this, means more people are using the other tools, and funding the less "moral" versions.

Turns out, people dont really care about Morals all that much, but they do like to argue about them.

I've seen some amazing stuff from midjourney lately. Maybe I will try it out tomorrow.

Yeah. Dall-E has a lot more limitations than Midjourney. Stable diffusion has even less.

The OpenAI crowd are worried about people using it as a weapon or to mess people around.
They are trying to be morally correct and all that. But the simple fact that they do this, means more people are using the other tools, and funding the less "moral" versions.

Turns out, people dont really care about Morals all that much, but they do like to argue about them.

Hi.

Quote from: MaNaeSWolf on September 28, 2022, 06:48:01 PM

Yeah. Dall-E has a lot more limitations than Midjourney. Stable diffusion has even less.

The OpenAI crowd are worried about people using it as a weapon or to mess people around.
They are trying to be morally correct and all that. But the simple fact that they do this, means more people are using the other tools, and funding the less "moral" versions.

Turns out, people dont really care about Morals all that much, but they do like to argue about them.

Hi.

This seems to be even more true today than it did when I wrote that not all that long ago.

Midjourney seems to be pulling a head of the pack, as they have revenue stream from selling services.
Stable Diffusion is where the technical guys are going. This is where you can have really fine control over your outputs.
Dall-e seems to be lagging in quality now.

I expect to see a lot more commercial options in the next few months. exciting times