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

Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

generrated.com

Post reply on HN
Print view View original

21–30 of 37 posts

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#21

Post by skybrian »

whywhywhywhy wrote: post #2

End of the day, unless it's opened up Dall-E 2 will be seen as an evolutionary dead end of this tech and a misstep. It's gone from potentially one of the most innovative companies on the horizon to a dead product now I can spin up equivalent tech on my own machine, hook into my workflow and tools in an afternoon all because Stable Diffusion released their model into the wild.

Casual users don't have a workflow to hook into, though. A website will be more convenient for them since there's nothing to install, and the web app probably runs faster than whatever hardware they're using.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#22

Post by tiluha »

jimmygrapes wrote: post #16

The best thing about this imo is that it proved to me that these image generation algorithms aren't just regurgitating minor variations on an existing image somewhere in their database of billions of images. Although I understand the general idea of how these work, I still had my doubts. The astronaut one in particular made this stand out; many of these artists definitely did not draw astronauts, yet the resultant im…

I actually found 'the discovery of gravity' far more interesting. Depending on the style the interpretation was completely different.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#23

Post by capitalsigma »

filoleg wrote: post #15
capitalsigma wrote: post #12

Earlier quoted context omitted.

Here, a less specific prompt: "Portrait of Joe Biden in the oval office, 4k render" First attempt: https://pasteboard.co/IYo5m6KeaqF4.png OK, I'll grant you that all of the parts of a face are there and reasonably correct (except for the bottomless pits of darkness in his nose and mouth) Second attempt, I end up with these weird artifacts in his head half of the time (3/5 of my generated images) * https://pasteboard.…

What is your guidance scale number, the number of iterations, and the chosen sampler? Those would be very relevant to know. Pretty much the most relevant thing aside from the prompt itself. Setting guidance scale number higher typically results in imagery getting trippier and more surreal with more artifacts. So i feel like that's the main culprit for the artifacts. I am pretty curious to see how far we can get with…

I'm using the default settings on the webui, here are the parameters:

``` Portrait of Joe Biden in the oval office, 4k render seed:1331361607 width:512 height:512 steps:50 cfg_scale:7.5 sampler:k_lms ```

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#24

Post by nextaccountic »

whywhywhywhy wrote: post #2

End of the day, unless it's opened up Dall-E 2 will be seen as an evolutionary dead end of this tech and a misstep. It's gone from potentially one of the most innovative companies on the horizon to a dead product now I can spin up equivalent tech on my own machine, hook into my workflow and tools in an afternoon all because Stable Diffusion released their model into the wild.

I wish the same happened with GPT-3

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#25

Post by filoleg »

capitalsigma wrote: post #23
filoleg wrote: post #15

Earlier quoted context omitted.

What is your guidance scale number, the number of iterations, and the chosen sampler? Those would be very relevant to know. Pretty much the most relevant thing aside from the prompt itself. Setting guidance scale number higher typically results in imagery getting trippier and more surreal with more artifacts. So i feel like that's the main culprit for the artifacts. I am pretty curious to see how far we can get with…

I'm using the default settings on the webui, here are the parameters: ``` Portrait of Joe Biden in the oval office, 4k render seed:1331361607 width:512 height:512 steps:50 cfg_scale:7.5 sampler:k_lms ```

Thanks for providing the seed, because that would let me show you how exactly the parameters can affect your specific image without generating a "random" different one every time.

Check out the exact same seed and prompt and cfg_scale, but with the steps aka iteration number at 100 (50 in general feels way too low, even for the samplers that are kinda good with low iteration numbers).

https://pasteboard.co/I6yXg5mZip6D.png

Obvious glitchiness in the face. Below is the same one, but with a k_euler_a sampler (I don't use k_lms, mostly k_euler_a or k_dpm_2_a) + 100 iterations.

https://pasteboard.co/xaTJiN6eVhm2.png

Less glitchiness, but Joe looks more caricature-like, than real. And also, not super quite like Joe. Let's try the same, but at 150 iterations and set the CFG at 10.

https://pasteboard.co/a9OigPXS9Ky1.png

Not much different in terms of realism, but the person looks distinctly way more like Joe. Let's up the iteration number to 200.

https://pasteboard.co/ey0ZzC110CrK.png

We got a bit closer to what we wanted. Faces are a bit of a difficult thing to do, but i think we can figure it out. Overall, it feels a bit "wobbly". I noticed that it tends to be beneficial to decrease the CFG as you increase iteration number, if you want photos to be more photorealistic. Let's set it to 6, and the iteration at 200.

https://pasteboard.co/na1fH54LkqO2.png

I would say this looks pretty good, but I think we can do better. The important part is imo the prompt, and I think we can edit yours to get a bit better results. Here is the result for "portrait of Joe Biden in oval office, dslr" with 200 iterations, CFG at 6, and k_euler_a sampler.

https://pasteboard.co/noHbMgxhLU4s.png

That one was probably my favorite (or maybe it was the one before).

Overall, you can play with this almost infinitely. Adding different words to the prompt in different spots can yield pretty different results. And that's not even mentioning all the parameters one can tune.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#26

Post by davidbarker »

glial wrote: post #3

DALL-E tech aside, this is a great reference for artistic style.

Thank you! I actually ended up learning far more about particular artists than I expected. Either names I'd heard of but didn't know much about (https://generrated.com/prompts/edvardMunch), or brand new artists whose style I now love (https://generrated.com/prompts/okudaSanMiguel).

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#27

Post by davidbarker »

ik8s wrote: post #5

Thank you for sharing this! I realized it was really hard to verbalize what I wanted from Stable Diffusion, so will be trying some of these and see what it comes up with.

Wonderful to hear — I really hoped it would provide some inspiration. It's great that our prompts can be (almost) anything, but it also makes it somewhat overwhelming to know where to start. I hope these can work as some starting points.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#28

Post by davidbarker »

lukeplato wrote: post #20

How much did it cost?

In total, I've spent around $500 with DALL•E 2 but I would say only half of that created images that went into the site.

Many of the prompts gave great results on the first try, but some of them required 5/10 attempts to get the prompt just right.

I've saved all of the generations I've created and at some point I'd love to upload them all to show how slightly different text can really change the images generated.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#29

Post by davidbarker »

jimmygrapes wrote: post #16

The best thing about this imo is that it proved to me that these image generation algorithms aren't just regurgitating minor variations on an existing image somewhere in their database of billions of images. Although I understand the general idea of how these work, I still had my doubts. The astronaut one in particular made this stand out; many of these artists definitely did not draw astronauts, yet the resultant im…

Yes! This is one of the most interesting aspects that I found. If it knows the particular style, it's typically very good at knowing how to use it.

I'm glad it could help clear things up for you.

Re: Show HN: I made 7k images with DALL-E 2 to create a reference/inspiration table

#30

Post by davidbarker »

tiluha wrote: post #22
jimmygrapes wrote: post #16

The best thing about this imo is that it proved to me that these image generation algorithms aren't just regurgitating minor variations on an existing image somewhere in their database of billions of images. Although I understand the general idea of how these work, I still had my doubts. The astronaut one in particular made this stand out; many of these artists definitely did not draw astronauts, yet the resultant im…

I actually found 'the discovery of gravity' far more interesting. Depending on the style the interpretation was completely different.

I'm glad you noticed this too.

With "a horse", you always get a horse, but with something more abstract like "the discovery of gravity" or "a representation of anxiety", it's far more variable between a scene — perhaps people performing an experiment, or someone falling out of a window(!) — or it's simply a large block of text without any people in the image.

Post reply on HN