Hacker News Re-Imagined

Nvidia releases Paint me Picture – A web app for GauGAN2

  • 161 points
  • 1 day ago

  • @nitinreddy88
  • Created a post
  • • 87 comments

Nvidia releases Paint me Picture – A web app for GauGAN2


@King-Aaron 21 hours

Replying to @nitinreddy88 🎙

Trying to do everything that the tutorial and other commenters have suggested, but when I click the arrow button it just waits for a few moments and nothing is generated. Am I missing something?

Reply


@togaen 11 hours


@thuccess129 20 hours

Nvidia should visualize a periodic 24 hour 3D landscape sweep of the chatter on social media platform for metaverse dive through and interactive engagement.

Reply


@nitred 18 hours

I wonder if in a decade or so, large tech companies will unseat Disney, Warner Brothers etc as creators of animation movies.

While the results on Nvidia's website aren't too impressive, if you look at the history of animated movies [1], one can see how trivial and simplistic the art and animation was.

Having had some experience doing some research on GANs at university, I know them to be very powerful. What's very important to note is that the images generated my the model are truly "novel" i.e. completely fictitious. The images generated may be biased to some of the training data such as color and texture of the water and rocks for example, but every image is a fantasy of the model. The only way the model can generate such realistic images is because it has a very good abstract internal representation of what oceans, waves, rocks are.

Back at university, I pitched the idea to my professor of using GANs for generating "novel" images in real time while parents would read bed time stories to children. I didn't get very far. Glad to see some real progress in that direction.

[1] https://www.filmsite.org/animatedfilms.html

Reply


@Tade0 16 hours

MtG dual-color lands serve as a good source of ideas on what to put in the textbox:

https://www.mtglands.com/coloridentity-dualcolor.html

As for the UI: layout via tables?

Reply


@system2 20 hours

After trying 30 minutes I kind out understood some stuff. But the ui is legit anxiety inducer. I hope they can fix the ui to make it fun. Currently felt like using 80's DOS graphic software with so much manual input.

Reply


@stefan_ 1 day

This seems to be the link: http://gaugan.org/gaugan2/

Reply


@jeroenhd 14 hours

The UI for the demo is atrocious, but that's probably because the text-to-image generation was glued to their existing AI painting tool.

I'd love for just the algorithm generation tool to be available for download. The web UI is clunky and just doesn't seem to work right.

Reply


@poniko 19 hours

The UI is horribel .. they don't have one single ux person who have 2 hour to spare that could help out at nvidia ...

Reply


@blackoil 20 hours

I am impressed, obviously still a tech. demo but I like the future. text to image isn't awesome but segmentation worked beautifully

https://imgur.com/a/VSv9ZbA

Reply


@akersten 23 hours

I really don't have anything constructive to say. I think in general we're getting too soft on shitty things, so I'm going to be harsh.

I clicked through to the demo site ( http://gaugan.org/gaugan2/ ) and it was horrible.

The interface is clunky, slow, and confusing. I actually had to zoom out in my browser to see the whole thing. Had to click through a non-HTTPs warning. The onboarding tutorial is pretty bad.

I got a generic picture of the milky way for any prompt I tried ("rocks", "trees"). If you press Enter in the prompt field it refreshes the page.

This feels like a hackathon front-end hooked up to an intro to PyTorch webservice. It's only neat because, unlike the other 20 copies of this same project I've seen, it was the only one that didn't immediately throw its hands up and say "server overloaded, please wait."

If I'm meant to be impressed or fearful of "big data deepfake AI," this isn't it.

[0] https://imgur.com/BNLDt6A

Reply


@monkeydust 11 hours

Wow. I have a good degree of respect for Nvidia but this should never have been released in the state it is. Whos the product manager for this?

Reply


@AbuAssar 23 hours


@stunt 7 hours

It seems that the web UI was generated by AI too, because it's really hard to make sense of it.

Reply


@ribit 19 hours

Are there any open-source models that can do similar type of landscape generation? I would really like to look at the code and try to understand how these things are built...

Reply


@greenseagull21 5 hours

The comments are overwhelmingly critical of the user interface, which is undoubtedly the weak part of this release, but I was still able to get some very impressive results.

An AI generated house on a lake: https://imgur.com/a/0wtVKum

I have found the best results come from uploading an image, then using the demo tools to get a segmentation map and sketch lines, then editing those as you desire. Changing the styling at the end also makes a big difference!

Reply


@Baeocystin 20 hours

I like what it did with colorless green ideas sleeping furiously. It fits the mood of the sentence.

https://i.imgur.com/jPd0QLE.png

Reply


@saberience 21 hours

I entered "mountains" and "mountains and lake" and only got pictures of what looked like blurred galaxies and stars. Clicking any of the style buttons got me colored/tinted pictures of stars. Is it broken?

https://imgur.com/a/SiQwBA2

Reply


@Havoc 16 hours

What license are the produced images under? I could see this being used for cheap stock photos

Reply


@h2odragon 13 hours

I hear they have a miraculous new AI tool that magically determines your sexual desires then uses lasers to induce those feelings through your eyeballs with no contact necessary! Coincidentally demonstrated at this very same URL!

Even then, I don't think I'd care enough to fight through the layers of bullshit here.

Reply


@eis 1 day

I am just getting very weird results that don't look at all like the one in the demo video. Here for example the image it gave me for "car in front of house" https://i.imgur.com/QdtrtCR.png

Or how about this one for "dog playing with ball" https://i.imgur.com/ldGLdwF.png

I have tried about a dozen different input phrases and every time I get these very strange results.

Reply


@blt 1 day

I entered "kitten" and got typical surreal GAN output with disconnected topology, dozens of eyes, etc.

Edit: Looks like it was only trained on landscape images.

Reply


@edumucelli 18 hours

I input "a horse", it gives me this: https://imgur.com/1coGQix

Reply


@bradhensen 13 hours

Now, it's not an imagination if you can create a visual art using texts.

Reply


To whomever came up with this name, good job

Reply


@bliss 11 hours

I sit here quite impressed with my Pseudo 50s SciFi book cover

https://imgur.com/a/NZuGTXU

However... This gallery on imgur gives a better idea of capability https://imgur.com/gallery/coWN44P

Reply


@carrolldunham 23 hours

I finally got through to the demo through three links and it's so busted in so many ways for me that I give up. Maybe it's stupid to try with my old netbook but I don't get any indication of whether I need a fancy graphics card for it to work or if it's running on my end. Anyway

-the screen zooms around disorienting for the tutorial and I get to congratulations you made your first image - there's nothing there.

-Exiting the tutorial, checking 'text' instead of 'segmentation' just immediately switches back after entry

-The whole site is a fixed width which is wider than my screen

- A red alert check box at the bottom confuses me about whether that's why it's not working etc

Reply


About Us

site design / logo © 2021 Box Piper