# New AI Research Solved The Problem Photoshop Never Could!

## Метаданные

- **Канал:** Two Minute Papers
- **YouTube:** https://www.youtube.com/watch?v=Ab9gJv-lrOw
- **Дата:** 14.08.2025
- **Длительность:** 6:43
- **Просмотры:** 59,792

## Описание

❤️ Check out Lambda here and sign up for their GPU Cloud: https://lambda.ai/papers

Guide:
Rent one of their GPU's with over 16GB of VRAM
Open a terminal
Just get Ollama with this command - https://ollama.com/download/linux
Then run ollama run gpt-oss:120b - https://ollama.com/library/gpt-oss:120b

📝 The paper "Physically Controllable Relighting of Photographs" is available here:
https://yaksoy.github.io/PhysicalRelighting/

📝 My paper on simulations that look almost like reality is available for free here:
https://rdcu.be/cWPfD 

Or this is the orig. Nature Physics link with clickable citations:
https://www.nature.com/articles/s41567-022-01788-5

🙏 We would like to thank our generous Patreon supporters who make Two Minute Papers possible:
Benji Rabhan, B Shang, Christian Ahlin, Gordon Child, John Le, Juan Benet, Kyle Davis, Loyal Alchemist, Lukas Biewald, Michael Tedder, Owen Skarpness, Richard Sundvall, Steef, Sven Pfiffner, Taras Bobrovytsky, Thomas Krcmar, Tybie Fitzhugh, Ueli Gallizzi
If you wish to appear here or pick up other perks, click here: https://www.patreon.com/TwoMinutePapers

My research: https://cg.tuwien.ac.at/~zsolnai/
X/Twitter: https://twitter.com/twominutepapers
Thumbnail design: Felícia Zsolnai-Fehér - http://felicia.hu

## Содержание

### [0:00](https://www.youtube.com/watch?v=Ab9gJv-lrOw) Segment 1 (00:00 - 05:00)

This is a new AI relighting technique that almost no one's heard of and it's a bit like recreating the Mona Lisa in a laser lit club night. So, we talk about 3D modeling programs like Blender here where you can create a 3D scene and change the lighting with just one click. But that is not the case for real 2D photographs. For these, we load up Photoshop and even today in the age of AI, we can't easily move a light source. No, we can only do these pitiful edits to it, like changing the contrast. By the way, this is not me. I can only wish I could rock a fabulous mustache like that. So, the old rule still stands. The light is permanent until now. This amazing new paper promises a new way to relight a 2D photograph. Let's take this input photo. Remember, not a 3D scene, just a 2D photo. And now hold on to your papers, fellow scholars, for changing the time of day. My goodness, that looks absolutely incredible. Or just throw the lights out and make a cool nighttime scene with a new light source. So good. But how is that even possible? Well, step number one, to relight the scene, first we have to delight. Remove the lighting. No worries, that can be done with a previous paper. Easy peasy. Now what? of course, load it into Blender. And wait a second, this is still just a 2D image, not a 3D scene. So now, step number two, let's 3Dify it. Once again, using the power of the papers. Yes, we finally get a 3D scene. Hooray. And a newer problem. I mean, look, this is 3D. All right. But it does not look detailed at all. It even has holes in it. Let's give it a try. But render it with a new environment and hello. The inaccurate 3D information is clear as day. So we are kind of stuck. What do we do now? Dear fellow scholars, this is two minute papers with Dr. Carola Eher. Dr. Carol John Alpha. Well, let's look at the good part. At least it's now a 3D scene. So it's a little virtual world, so new light sources can be added. Okay, so now step number three. Let's beautify this scene somehow. And this is where the new paper comes in. The goal would be that this rough image would be rerendered into a beautiful photo. Let's call it a neural renderer where the input is a rough rendering and the output is a real image it corresponds to. Of course, to be able to train a neural network like that, we would need thousands of these pairs to train it on. Okay, if we look at the right part, this is trivial. The internet is full of photos, but what about rough 3D renderings that correspond to these photos? Well, of course, that's the hard part. Earlier, we said that we can produce 3D scenes from photos, but there is one problem. This is still not a perfect match. Why? Let's see. Aha. Because the lighting is different. We know the scene from the photo, but we don't yet know the kind of lighting that was used to produce it. For instance, the lights here aren't really lights. So, here's the idea. Let's throw a few lights in there, render the scene, and see how different the result is from the target photo. H well, at this point, it is not close at all. Now, we wiggle them around a bit and see if it is closer. Repeat this process for thousands and thousands of photos until the 3D scene with the lights really explains the photo. And now you have a proper neural network for neuro rendering. And this is the key contribution here. Putting all of these together, you can now finally take a photo and change it from night to daytime or the other way around, remove lights or add lights. Goodness. Look, they even cast shadows. And spotlights and area lights also work. But it gets better. Even animated projectors can be used so you can put yourself in the craziest ways onto any photo now. And this is why it's a bit like the Mona Lisa at the laser lit party. What a time to be alive. So, does this take hours to run per photo? Well, here is the best part. Not at all. All this runs within three seconds. Two seconds for pre-processing and relighting less than a second. Crazy. Wow. And unfortunately, it breaks my heart that it flew under the radar. I hear almost nobody talking about this paper. Insanity. Now, before we proceed to limitations, please watch this video until the end and leave a kind comment.

### [5:00](https://www.youtube.com/watch?v=Ab9gJv-lrOw&t=300s) Segment 2 (05:00 - 06:00)

Otherwise, it is really hard to get the YouTube algorithm to recommend this kind of content. Now, the results are not perfect. You see that as the lighting is moving around, things are a bit blocky. I think the resolution of the 3D geometry is much better than previously, but still not fantastic. Also, if you place lights behind the scene or at some unexpected places, you will also get artifacts. Also, images with lots of specular highlights and complex materials like skin are still a no-go. Too hard for now. But just think about what we will be capable of two more papers down the line. My goodness, today we are witnessing the fundamental transformation of the photograph from a static memory into a living editable world. This isn't just about making cool edits. It's about handing the artist the power to direct reality itself. long after the shutter has closed. And it is absolutely brilliant. This paper and the quality of presentation they do in their video needs an award. Full stop. And I got to say, I can't stop playing with OpenAI's Open GPT model through Lambda GPU cloud. And as you see, I am doing very useful things with it for science. Yes, this is actual speed. I can't believe that I can have more than 100 billion parameters running super fast here. Many of you fellow scholars are using it and if you don't, make sure to check it out. It costs only a couple dollars per hour. Insanity. You can rent an Nvidia GPU through lambda. ai/papers AI/papers or click the link in the video description.

---
*Источник: https://ekstraktznaniy.ru/video/12177*