OpenAI's DALL-E 2 Has Insane Capabilities! 🤖

Use the code TWOMINUTE at checkout to get 10% off!

📝 The paper "Hierarchical Text-Conditional Image Generation with CLIP Latents" is available here:

☀️My free Master-level light transport course is available here:

📝 Our Separable Subsurface Scattering paper with Activition-Blizzard:

📝 Our earlier paper with the caustics:

Reynante Martinez, the master's page:

Rendered images:

Hotel scene:

Path tracing links on Shadertoy:

Caustics:

Dispersion:

Chapters:
0:00 Teaser
0:48 Light Transport
1:18 Variant generation
1:48 Experiment 1
2:20 Let's try it again!
3:40 Experiment 2
5:05 Experiment 3
6:34 Experiment 4
7:40 Indirect Illumination, dispersion, course

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Aleksandr Mashrabov, Alex Balfanz, Alex Haro, Andrew Melnychuk, Benji Rabhan, Bryan Learn, B Shang, Christian Ahlin, Eric Martel, Geronimo Moralez, Gordon Child, Jace O'Brien, Jack Lukic, John Le, Jonas, Jonathan, Kenneth Davis, Klaus Busse, Kyle Davis, Lorin Atzberger, Lukas Biewald, Luke Dominique Warner, Matthew Allen Fisher, Matthew Valle, Michael Albrecht, Michael Tedder, Nevin Spoljaric, Nikhil Velpanur, Owen Campbell-Moore, Owen Skarpness, Rajarshi Nigam, Ramsey Elbasheer, Steef, Taras Bobrovytsky, Ted Johnson, Thomas Krcmar, Timothy Sum Hon Mun, Torsten Reil, Tybie Fitzhugh, Ueli Gallizzi.

Károly Zsolnai-Fehér's links:

#openai #dalle

[TwoMinutePapers]

Use the code TWOMINUTE at checkout to get 10% off!

[truemori]

I can't wait for caustics to become able to be processed in games and such

[basicnpcc]

Two Minute Papers: 'Are you thinking what I'm thinking!?!'
Me: 'Chipmunks with capes?'
Two Minute Papers: 'Denoising!'

[Soporonix]

Great video, you should do a video on midjourney's new update, it seems to be even better than dall e 2 in some aspects and even with simple prompts its really good

[scrpin]

You're trying to get DALL-E 2 to do what it's not designed to do. The variations you see here are in fact coming from the training data composited cleverly together from your text prompt. It does not compute the caustics. The caustics are there in the training images. The subsurface scattering is also there in the training data. The model understands what subsurface scattering looks like so it's just compositing the training data together and applying the SS style on it to make it match closest to the trained style, not that it computes anything.

[James-iptc]

I was conducting research on stable diffusion and severe caustics, and I obtained some pretty intriguing findings. It is my hope that they will modify the training such that you will be able to explore with light and refraction in ways such as animation with caustics that would take a machine an amazing amount of time to manually compute.

[alkenstein]

Since DALL-E 2 and Stable Diffusion both work by starting with noise and resolving until the noise has disappeared, I was expecting that you were going to present these algorithms with a noisy light transport sim and they would take over to resolve the outcome much more quickly. No idea if that could work but it sounds reasonable?

[lp]

Is your "What a time to be alive!" sample part of your cc audio library?

[pandoraeeris]

Get an AI to generate lyrics and music to the song, "Two More Papers Down The Line".

[DonnaPinciot]

4:57 "We don't look at the drink"
That's right! We look at where we will be 2 more drinks down the line!

[lugui]

I've been using Dalle for quite some time now, I found that making 3d scenes is a good way to drive the variation generation to get the style, camera angles and lighting of the scene

[aimanifest]

Very interesting interpretations and analysis! What a great time to be alive🎉

[_koy_]

It is always going to be a great time when you upload. Thanks for making entertainment that somehow is Educational.

[garethbridges]

You should use an AI voice synthesis for one of these episodes and see how many of us can tell :)

[Lumynex]

the video is really cool! and the costics are amazing as always. could you try doing this with the Midjourney ai version 1.4 though? that came out a week ago and lots and lots of people are saying that it produces way better images than dall e 2. the previous versions of midjourney pretty bad in comparison to dall e or stable diffusion, but the new version seems to beat them all!

[I_Was_Named_This_Way...]

I don’t read papers, I watch your videos 😉

[DiegoAlanTorres]

Dall-E 2 looks so ugly to me now, especially after seeing Stable Diffusion and Midjourney V4 now.
You'll have to practice catching up in the future with how fast things are going.

[AcidZero]

Where can I find the picture of the ring casting a heart shadow? Thanks

[jenkem]

Holy crap, amazing experiment! So interesting.

[edd]

I would be interesting to see if stable diffusion is able to de noise images as it gives more control