@@kingx1180 eh, AMD and Nvidia are two very different companies. Nvidia no doubt has many of the best engineers, but AMD at least hasn't forsaken the gaming sphere in favor of the AI/compute industry. Sure, AMD wants a piece of that pie, and they're trouncing intel on the CPU side of things, but Nvidia has seen most of the explosive growth in the compute sector. To the point that it makes them more money than consumer graphics cards. Couple that with the past few years of crypto mining and PC Gamers are stuck with choosing between capable yet overpriced and, well, abused hand-me-downs. Nvidia is no longer a graphics card company they are an AI hardware/software company.
@@jackinthebox301 Are you actually gonna compare the magnitude of these industries? It sure is more realistic than Meta's new media from social network into VR universe. (Please read this in a calm way, I'm just objective, thank you).
@@v.f.38 I honestly have no idea what you're trying to say. It's been a minute since I looked at their financials, but if I remember correctly commercial compute surpassed consumer graphics revenue for Nvidia in 2020.
I've been waiting for exactly this, for years. I dabble in permaculture... basically landscape design for the purposes of enhancing the utility of the land. One thing you do often is "shadow mapping"... if you plant 300 different species of plants, you have to factor in time in their placement, as some plants will grow to shade others, slowing their growth. So you need to accurately predict how things will look as the sun moves across the sky, from sunrise to sunset, across seasons, and over the course of years or even decades. I know where the shadows of a building will fall because the building doesn't move and grow. But plants do, and at variable rates which are dependent on the growth of their neighbors. You can very quickly end up with millions of variables. This can actually increase things like crop yields or improve wildlife habitat if incorporated into planning. It's not just about pretty pictures, you can feed people with this technology. There are tons of applications for this. It could reduce home heating and cooling energy costs. It can significantly reduce water usage and fertility loss on land. In the hands of experienced planners, this could enhance windbreaks and reduce soil erosion. And I can easily imagine similar impacts on civil engineering and city design. This could impact commodity futures and influence financial markets. I understand this is not designed as a predictive model, but it's a huge step in that direction, and the most difficult one to crack. We've used similar methods, much more crudely for decades with far poorer results and still optimized various design systems by astronomical margins. This could be a game-changer in ways far exceeding the intent of the developers. Current methods basically involve scanning in models of hundreds of plants per species for hundreds of species and loading them into something like UnrealEngine to do the lighting prediction, and then swapping models of each plant over time to one of a later growth stage under the respective conditions of the original model to create virtual time-lapses. It's massively labor-intensive and often cost prohibitive. This is about 70% of the way to replacing those methods, and very cheap on resources. This is why I like nVidia's research. It has applications outside of simply making images for the sake of making images. It has real-world applications.
I bet there are models specifically designed to predict plant development and maybe even they can be integrated into game engines, or Blender, to produce lighting simulations. This system only smoothens the transitions between video moments, selectively keeping or discarding visual data. Probably not very useful for your use case. Maybe with some fine tune someone could adapt this to simulates growth of plants along with their shadows, maybe, but it's a log shot.
I'm not sure I understand what value this adds. I mean, this isn't exploring real changes, it's just trying to make intermediate steps between real photos that look "normal". Not only is this not a predictive model, it isn't a plant growth model at all. It's a plant "appearance" model. Unless you already have a picture on both ends of the time scale you're interested in, this model gives you nothing, and when you do, this model gives you something that looks like the plant in question, but only to the extent that it's moving the image from point A to point B. Nothing about the technique here implies the ability to extrapolate beyond the endpoints, at all, even hypothetically, and even if it did, it wouldn't be using any real data about the plant and couldn't predict anything real about its growth.
@@michaelleue7594 That's a fair point, to clarify, we do in fact have long duration timelapses. It's specifically the dynamic lighting that is of interest. It's impractical to figure out the lighting at every stage of development in different locations. The angle of the sun in the sky at various times of day and night, throughout the season. The fact that it isn't 3d doesn't matter if you can take the pixel data from a simulated light source. You could have 20,000 images taken in timelapse of a single plant, but under largely artificial lighting. How the light hits it relative to it's surroundings at a different time / date / longitude and latitude however is more complicated. Because it can smooth lighting and apply it to a scene taken under different lighting, you can then use a still image and get the estimated luminosity of the pixel data, dump those values to a spreadsheet and use linear regression on the table of pixel luminosity data to work it out for a volume estimate. The predicted numbers can be used as an efficiency rank. It just needs to look similar to what the actual scene would be under different lighting with a fair degree of accuracy. It's definitely using outside of it's intended purpose. But I think it could be adapted to suit that need better than the tools we currently have, specifically because it can work from timelapse images, and doesn't need a volumetric model of each stage of growth.
This is so cool. Since this is open source I want to try this on one of those human timelapse videos with titles like "I took a picture every day from age 10 to 20". I will try it this weekend, hopefully there is one of such videos under open source license so I can post it.
I'm gonna try this on my timelapse images, normally professionals use camera techniques and post processing to get the desired effects but this could reduce lot of hard work that is required to get decent looking simple timelapse videos & especially help if you were not able to capture enough images on the spot
3 years ago, I took pictures of growing crops from different angles, every day at 10am, during 3 months. Due to the lack of consistency of the framing and weather conditions, I had almost given up trying to make something out of it. A new hope just arised! Thanks Karoly and Nvidia :)
"How on Earth do we express these videos as pictures in the paper?" In my opinion, missed opportunity by the researchers to turn their research paper into a flipbook.
To be honest I feel when we get to the stage where we can make everything perfect (in videos/images, art and the like) and change or create everything instantly on a whim, then it starts to lose its meaning. Paraxodically, at that point it is the imperfections that become meaningful.
That's a critical view of things, but meaningless in a broader sense. Whether or not the art or otherwise these AI effect in whatever ways is meaningful is really only a personal, subjective question. These are powerful tools, and therefore you can expect them to take their place where appropriate and allowed across the entire market. If that doesn't happen, then we're mistake on their potential and power. What happens to AI on the market across society is where the discussion becomes more interesting to me, as that takes it to the level of the actual effect across billions of people; society always has to be careful but open in the management of powerful tools, least they give the public nukes but also not so tyrannical that we'd make driving cars illegal for example. Point being: the discussion of how it is appropriate to use AI should come before the discussion of how meaningful the works it touches/generates are.
I wonder if images and videos should remain admissible in court, seeing the amount of realistic tempering made possible by the AI research in the recent years…
Most concerning thing about the whole Ai mangling with graphics is the bending and distortion of video material to let visual reality show a circumstance that never happened. I get it, it’s solely to be able to adjust and tweak certain details to a less glitchy outcome, that’s the noble intent and I know It comes with the process and is in the nature of the development to have the most amount of control of the outcome but this aspect is certainly on the minds of the powers that be to be used against us and I can imagine a ton of morally wrong examples this stuff will be heavily exploited to keep certain info from us in the future as if censorship isn’t already rampant. Crazy times…
People are naive. We are playing with fire. No restrictions. The impact of mobile phones and internet will pale in comparison. And we are still struggling and do not understand that impact fully. The tech guys are dreaming up an utopia. Let's hope it won't turn into a dystopian, orwellian nightmare.
While this is interesting and I love it, the more I learn about AI generated/aided images, and the more these AI advance, the more i am worried about our ability to distinguish truth from fiction. We seriously need to thing about useful ways to give everybody the tools necessary to understand the difference.
Totally agree. The more I saw these AI technique come out for fooling human eyes , the more I wish I could live a simpler life and stay away from any digital things.
To notice the difference. At some point there won't be any way to tell real or made up images apart if there is no visible watermark on the generated images/videos. Add to that the arrival of fully immerse VR and voila! People won't care about reality anymore if it is just as fake as VR. And slowly we will descent into a world were nothing will matter to anyone except fast and small moments of empty gratification through a connected device.
@@rkan2 A blockchain does nothing to address the problem, it's as useless as writing your own certificate of authenticity for a photo you edited in Photoshop. You're not certifying nothing has been changed in the image, you're just saying you were involved in some way (possibly only uploading the image).
Maybe that's why they're moving. The AI mistook the chairs for animals so it allowed them to move like the plants. And since they are animals, the only acceptable movement patterns for them through time was that of an animal.
I wonder how this would work in an environment full of animals, for example a pond timelapse. As it seems like schools of fish might cause issues, a bit like those chairs. Oneday when i get a new gpu I'll test this.
I think the next big careers will be primarily in scientific data collection for AI systems to train on. If so, this implies new sensory technologies for high sensitivity data capture and accuracy. This might even lead to mobile devices being upgraded with additional sensors to increase data capturing quality and capacity; ultra violet light, temperature maps and 3d lidar scans of objects and the environment etc. Obviously, I'm just speculating but I am curious about how training data will be collected and what technologies and careers will be created along the way.
An AI can probably do whatever you are suggesting. The next big careers will be politicians making laws regulating the use of AI because protests against the use of AI are on the way.
With an AI for just about anything, one wonders what the future lies for humans. We are creating a society based on computer generated text, images, music and pretty soon ideas. While advancements like this deserves recognition, I cannot help but feel that a dystopian future is on its way.
Yeah... its impossible to not imagine that. Optimistic view: It will make socialism/comunism finally viable and maybe even needed in a world where humans don't need to work. In a world where machines can do every work better than us, governments will have to adapt to this new reality and build a new system for us. It could be paradise, with better health care, food for all, less resource waste, etc, but... it could go bad in the end due to mental health. Pessimistic view: AI will make work extremely scarce, where only the elite survives in the market. Governments will not adapt well and a huge crysis will arise, with poverty everywhere, crime rising and war for countless reasons. It could also be the result of a perfect life created by the "optimistic view", where humans cannot deal with the lack of goals and challenges and mental health start to decline really fast. Imo, Ai should always be limited. Always a tool.
There is no way this guys voice is a real human voice. He's like a cartoon voice. I love it though. I want to hang out with this guy in a pub and tell each other stories of Chat GPT and Midjourney for a few hours lol.
If it weren't for the missing data, I'd approach this problem from a different angle: since movement is often sudden with lots of "dead space" in between, I would attempt to craft a temporal seam carving algorithm to consolidate the frames where little is happening while preserving the gradual changes. Combine that with optical flow to compensate for things like camera movement, and you should have a much more stable base from which to further refine the output.
While installing this, has anyone else been having an error involving MSBuild.exe? I have the correct 2019 version installed but can't seem to get around it.
2 min papers becomes a 7 minute, then I would like to play it on speed 3.5, but Google still haven't AI for this. Amazing paper regardless of speed, though
🤔 BBC and other documentaries use artificial falsh light to capture plant growth in night. It is consistent. For what could be city timelapse used I dont know. I asume the breakthrough is that it understands the lighting changes, and as NVIDIA pushes its DLSS 3 that produce ugly artifacts in games, when adding in between frames for smoother resolution. But as PC reviewers pointed out, games are smoother with DLSS 3 but has a huge latency, as frames must be computed. So adding more computing for DLSS 3 would mean better quality but even worse latency.
To be able to generate images using AI with speech/text is amazing! Even simple words can create the most satisfying looking images especially with the progressing technology on graphics. I'm curious to try create one using Bluewillow, what do you guys think?
I’m thinking of four-dimensional pictures, where you can watch famous landmarks not just in the 3 dimensions of space, but also in the fourth dimension which is time. And all this would be made from the thousands of tourist photos that already exist.
Hi Dr KZF Kan you please make a video explaining how data is saved inside these great algorithms? For example what kind of data structures are used to by style transfer algorithms, and what members do these data structures have (like arrays of ints? arrays of floating points...) representing what aspect of the style? Thank you.
In his earlier videos he may have already explained what you're curious about. I know he has one on "how AI sees" or something similar to that. It should be full of hallucinogenic images of various things like cats and dogs. And in a way your question doesn't really make much sense, but that's okay. Neural nets don't have traditional programming data structures. You will hear jargon like latent images or latent data when talking about information stored within a trained AI model. This is the weights of the various node connections and their output biases. From these is where the information is stored. But by virtue of them being a net, there is no 1to1 of any individual values and what information they represent. It is a cumulative effect of multiple values together representing that information. And so any 1 individual value is responsible for multiple representations. Think of it as there are multiple layers of "images" laid out in various configurations, across different regions of the model, and they overlap without issue. As long as no 2 "images" are identical you can store as many of such permutations as the size of the model can allow. There are many issues with this explanation of course, but hopefully this better illustrates how to visualize the "data structure" that can be found in AI models
Far from perfect -- it literally hurts my eyes to look at, because everything's got some weird ghosting, like I'm seeing double, and my eyes keep trying to adjust to it.
You know the future is eventually going to be great once this Technologies implemented for gaming. Imagine God of War, or Untitled Goose game 🤩😂 Thanks for the vid 😄
The day an IA will start to make better reviews of IA progress than this channel, I guess he will not longer said "what time to be alive!" In fact, the IA would be the one saying that :)
By the time an AI knows its a great time for it to be alive, maybe there won't be anyone left for it to talk to, except robots and other AIs: "What a time for us to be the only ones left alive!". 🙃
Seeing as AI could keep up with and promote it's own inventions, or follow feeds to see what other AIs are up to, this time may soon come. And rather than crediting teams of researchers, it may well be crediting itself.
@@perfectfutures yeah, these last decades our technology has grew up exponentially, but I guess this would be absolutely nothing compared to what is coming, we need to study over 40 years of specialization and experience to achieve (with lucky) one break thought in one specific field. One AGI can do it in all fields at once and then it can improve itself, first doubling power every year, then every 3 months, then few days until it knows everything that can be know.. I am here thinking in what I should learn next to improve my coding job.. But then I ask myself, what is the point? How much time we have left before IA can do everything better than us. We are the biggest bottle neck for any fast progress. IA developers are already finding all ways to skip that bottleneck.
AI is now at the forefront of technology as some like mobile phones are somewhat feeling stagnant nowadays just waiting for the next eureka moment. So far, i've been having a lot of fun using AI image generators like Bluewillow in testing out these AIs. Hoping for the best!
I would run the pre- and actual time laps still photos of the 1980 Mt. Saint Helen's eruption. There have been attempts to tween the still into "footage" augmented with CGI but they're HORRIBLE. This might do a much more interesting job.
Nvidia, Google, Meta, Microsoft, Open AI, Amazon, Apple. These companies are on AI race and i love it because the consumer will eventually be the winner!
Not really... they are working on AI to make more money off of people, they will keep increasing their subscription models, just to make sure that the consumer owns nothing and has to pay regularly for everything they use, alongside the increasing rent, not to mention inflation, food prices, etc... All these companies are evil and they deserve to fall.
@@cyancoyote7366 that’s true. But I think if they are many companies in the market, the prices will be competitive. Think about the monopoly that IBM had before Microsoft ,Apple , Dell etc were launched. The computer was very expensive back then. Competition is generally good for the consumer.
@@MasterKey2004 Apple doesn't have an history of showcasing faulty prototypes.They'll let others try and fail at AI UX before launching anything, just like they did with smartphones and vr headsets.
The results are very good, actually so good that many will accept results as reality; so, what is being developed to identify AI-produced results from reality?
hmm... i actually don't love the look of this. the details in detailed areas like leaves/branches/grass look like some false morph effect. smooth areas like clouds and fog look great though
It would be possible to teach a Neural based AI, to deconstruct the mind/thought patterns and reconstruct those patterns as an simulated virtual interactive image. This would be an Approximation. AI should be able to learn to translate our thought processes and it would be able to reconstruct these processes to manifest the reconstruction in a Virtual reality in realtime.
That's a little how some AI aided brain-machine interfaces work: you give the AI training data with both EEG readings and a description of what each person was thinking at the time of the reading and the AI learns what electrical patterns in the brain mean, so later it can literally read someone's mind (provided that the person is wearing a helmet full of electrodes 😂). There is no visual representation from that yet but probably once the adoption of brain implants such as Neuralink becomes massive, we will see huge advances in the interpretation, visualization and simulation of the human mind's processes. Maybe soon we will be able to explore each other's minds simply wearing VR lens.
Ok, I'm nasty, I know... "If only an AI exist... that... somehow... could make... your speech... smooth as well" 0:44 Sorry, I couldn't resist to make that joke 😂 and it's not meant to hurt your feelings
2 min papers about Nvidia are always satisfying…
Amd sucks lol
@@kingx1180 eh, AMD and Nvidia are two very different companies. Nvidia no doubt has many of the best engineers, but AMD at least hasn't forsaken the gaming sphere in favor of the AI/compute industry. Sure, AMD wants a piece of that pie, and they're trouncing intel on the CPU side of things, but Nvidia has seen most of the explosive growth in the compute sector. To the point that it makes them more money than consumer graphics cards. Couple that with the past few years of crypto mining and PC Gamers are stuck with choosing between capable yet overpriced and, well, abused hand-me-downs.
Nvidia is no longer a graphics card company they are an AI hardware/software company.
ikr, engineers at nvidia are amazing
@@jackinthebox301 Are you actually gonna compare the magnitude of these industries? It sure is more realistic than Meta's new media from social network into VR universe. (Please read this in a calm way, I'm just objective, thank you).
@@v.f.38 I honestly have no idea what you're trying to say. It's been a minute since I looked at their financials, but if I remember correctly commercial compute surpassed consumer graphics revenue for Nvidia in 2020.
I've been waiting for exactly this, for years.
I dabble in permaculture... basically landscape design for the purposes of enhancing the utility of the land. One thing you do often is "shadow mapping"... if you plant 300 different species of plants, you have to factor in time in their placement, as some plants will grow to shade others, slowing their growth. So you need to accurately predict how things will look as the sun moves across the sky, from sunrise to sunset, across seasons, and over the course of years or even decades. I know where the shadows of a building will fall because the building doesn't move and grow. But plants do, and at variable rates which are dependent on the growth of their neighbors. You can very quickly end up with millions of variables.
This can actually increase things like crop yields or improve wildlife habitat if incorporated into planning. It's not just about pretty pictures, you can feed people with this technology. There are tons of applications for this. It could reduce home heating and cooling energy costs. It can significantly reduce water usage and fertility loss on land. In the hands of experienced planners, this could enhance windbreaks and reduce soil erosion.
And I can easily imagine similar impacts on civil engineering and city design. This could impact commodity futures and influence financial markets. I understand this is not designed as a predictive model, but it's a huge step in that direction, and the most difficult one to crack. We've used similar methods, much more crudely for decades with far poorer results and still optimized various design systems by astronomical margins. This could be a game-changer in ways far exceeding the intent of the developers.
Current methods basically involve scanning in models of hundreds of plants per species for hundreds of species and loading them into something like UnrealEngine to do the lighting prediction, and then swapping models of each plant over time to one of a later growth stage under the respective conditions of the original model to create virtual time-lapses. It's massively labor-intensive and often cost prohibitive. This is about 70% of the way to replacing those methods, and very cheap on resources.
This is why I like nVidia's research. It has applications outside of simply making images for the sake of making images. It has real-world applications.
interesting read, thanks
I bet there are models specifically designed to predict plant development and maybe even they can be integrated into game engines, or Blender, to produce lighting simulations. This system only smoothens the transitions between video moments, selectively keeping or discarding visual data.
Probably not very useful for your use case. Maybe with some fine tune someone could adapt this to simulates growth of plants along with their shadows, maybe, but it's a log shot.
I'm not sure I understand what value this adds. I mean, this isn't exploring real changes, it's just trying to make intermediate steps between real photos that look "normal". Not only is this not a predictive model, it isn't a plant growth model at all. It's a plant "appearance" model. Unless you already have a picture on both ends of the time scale you're interested in, this model gives you nothing, and when you do, this model gives you something that looks like the plant in question, but only to the extent that it's moving the image from point A to point B.
Nothing about the technique here implies the ability to extrapolate beyond the endpoints, at all, even hypothetically, and even if it did, it wouldn't be using any real data about the plant and couldn't predict anything real about its growth.
@@michaelleue7594 That's a fair point, to clarify, we do in fact have long duration timelapses. It's specifically the dynamic lighting that is of interest. It's impractical to figure out the lighting at every stage of development in different locations. The angle of the sun in the sky at various times of day and night, throughout the season. The fact that it isn't 3d doesn't matter if you can take the pixel data from a simulated light source.
You could have 20,000 images taken in timelapse of a single plant, but under largely artificial lighting. How the light hits it relative to it's surroundings at a different time / date / longitude and latitude however is more complicated. Because it can smooth lighting and apply it to a scene taken under different lighting, you can then use a still image and get the estimated luminosity of the pixel data, dump those values to a spreadsheet and use linear regression on the table of pixel luminosity data to work it out for a volume estimate. The predicted numbers can be used as an efficiency rank.
It just needs to look similar to what the actual scene would be under different lighting with a fair degree of accuracy.
It's definitely using outside of it's intended purpose. But I think it could be adapted to suit that need better than the tools we currently have, specifically because it can work from timelapse images, and doesn't need a volumetric model of each stage of growth.
is this chatgpt
Wonder if we could also apply this to street view imagery to get extremely smooth time lapses
This is a really great idea
Please Microsoft, start integrating all these new AI to your products, so Google would need to do it also.
not only that, but an actual interactive walkable street view with WASD key
We are in a time where fake looks better and at times more real, than reality itself.
This is so cool. Since this is open source I want to try this on one of those human timelapse videos with titles like "I took a picture every day from age 10 to 20". I will try it this weekend, hopefully there is one of such videos under open source license so I can post it.
I'm gonna try this on my timelapse images, normally professionals use camera techniques and post processing to get the desired effects but this could reduce lot of hard work that is required to get decent looking simple timelapse videos & especially help if you were not able to capture enough images on the spot
3 years ago, I took pictures of growing crops from different angles, every day at 10am, during 3 months. Due to the lack of consistency of the framing and weather conditions, I had almost given up trying to make something out of it.
A new hope just arised! Thanks Karoly and Nvidia :)
"How on Earth do we express these videos as pictures in the paper?" In my opinion, missed opportunity by the researchers to turn their research paper into a flipbook.
Every time i wake up and watch one of your videos you make me smile with the chrisma you add to these videos
You are too kind, thank you so much! 🙏
To be honest I feel when we get to the stage where we can make everything perfect (in videos/images, art and the like) and change or create everything instantly on a whim, then it starts to lose its meaning. Paraxodically, at that point it is the imperfections that become meaningful.
We can get everything now.
That's a critical view of things, but meaningless in a broader sense. Whether or not the art or otherwise these AI effect in whatever ways is meaningful is really only a personal, subjective question. These are powerful tools, and therefore you can expect them to take their place where appropriate and allowed across the entire market. If that doesn't happen, then we're mistake on their potential and power.
What happens to AI on the market across society is where the discussion becomes more interesting to me, as that takes it to the level of the actual effect across billions of people; society always has to be careful but open in the management of powerful tools, least they give the public nukes but also not so tyrannical that we'd make driving cars illegal for example. Point being: the discussion of how it is appropriate to use AI should come before the discussion of how meaningful the works it touches/generates are.
We'll go full circle then, back to handmade human art that has flaws.
I wonder if images and videos should remain admissible in court, seeing the amount of realistic tempering made possible by the AI research in the recent years…
Well, I have about 40000 images of a building in construction, I could definitely use this!
If humans can make incredible stuff like this, imagine how amazing the papers written by the new AI will be!
1000x more incredible if the day comes when ai can do that
Most concerning thing about the whole Ai mangling with graphics is the bending and distortion of video material to let visual reality show a circumstance that never happened. I get it, it’s solely to be able to adjust and tweak certain details to a less glitchy outcome, that’s the noble intent and I know It comes with the process and is in the nature of the development to have the most amount of control of the outcome but this aspect is certainly on the minds of the powers that be to be used against us and I can imagine a ton of morally wrong examples this stuff will be heavily exploited to keep certain info from us in the future as if censorship isn’t already rampant. Crazy times…
People are naive. We are playing with fire. No restrictions. The impact of mobile phones and internet will pale in comparison. And we are still struggling and do not understand that impact fully. The tech guys are dreaming up an utopia. Let's hope it won't turn into a dystopian, orwellian nightmare.
While this is interesting and I love it, the more I learn about AI generated/aided images, and the more these AI advance, the more i am worried about our ability to distinguish truth from fiction. We seriously need to thing about useful ways to give everybody the tools necessary to understand the difference.
Totally agree. The more I saw these AI technique come out for fooling human eyes , the more I wish I could live a simpler life and stay away from any digital things.
As the radicals say "Adapt or die." I totally agree with your thoughts though.
All certain information needs to be based on some blockchain. It is the only way you could fix it. Until it busts the crypto!
To notice the difference.
At some point there won't be any way to tell real or made up images apart if there is no visible watermark on the generated images/videos.
Add to that the arrival of fully immerse VR and voila! People won't care about reality anymore if it is just as fake as VR. And slowly we will descent into a world were nothing will matter to anyone except fast and small moments of empty gratification through a connected device.
@@rkan2 A blockchain does nothing to address the problem, it's as useless as writing your own certificate of authenticity for a photo you edited in Photoshop. You're not certifying nothing has been changed in the image, you're just saying you were involved in some way (possibly only uploading the image).
5:50 I wonder if it's an illusion but it seems that the legs of the chairs are moving one leg at a time, like it was a chair animal.
I can't tell if a chair animal is an alice in wonderland character or an SCP
I'm more like.. Why shouldn't the chairs be moving if they moved in the dataset.
Maybe that's why they're moving. The AI mistook the chairs for animals so it allowed them to move like the plants. And since they are animals, the only acceptable movement patterns for them through time was that of an animal.
I wonder how this would work in an environment full of animals, for example a pond timelapse. As it seems like schools of fish might cause issues, a bit like those chairs. Oneday when i get a new gpu I'll test this.
I think the next big careers will be primarily in scientific data collection for AI systems to train on. If so, this implies new sensory technologies for high sensitivity data capture and accuracy. This might even lead to mobile devices being upgraded with additional sensors to increase data capturing quality and capacity; ultra violet light, temperature maps and 3d lidar scans of objects and the environment etc.
Obviously, I'm just speculating but I am curious about how training data will be collected and what technologies and careers will be created along the way.
Sounds logical. Probably labeling things too. Manual labour is not dead yet I think.
An AI can probably do whatever you are suggesting. The next big careers will be politicians making laws regulating the use of AI because protests against the use of AI are on the way.
With an AI for just about anything, one wonders what the future lies for humans. We are creating a society based on computer generated text, images, music and pretty soon ideas. While advancements like this deserves recognition, I cannot help but feel that a dystopian future is on its way.
Yeah... its impossible to not imagine that.
Optimistic view: It will make socialism/comunism finally viable and maybe even needed in a world where humans don't need to work. In a world where machines can do every work better than us, governments will have to adapt to this new reality and build a new system for us. It could be paradise, with better health care, food for all, less resource waste, etc, but... it could go bad in the end due to mental health.
Pessimistic view: AI will make work extremely scarce, where only the elite survives in the market. Governments will not adapt well and a huge crysis will arise, with poverty everywhere, crime rising and war for countless reasons. It could also be the result of a perfect life created by the "optimistic view", where humans cannot deal with the lack of goals and challenges and mental health start to decline really fast.
Imo, Ai should always be limited. Always a tool.
There is no way this guys voice is a real human voice. He's like a cartoon voice. I love it though. I want to hang out with this guy in a pub and tell each other stories of Chat GPT and Midjourney for a few hours lol.
Just amazing. The speed of development is insane!
is there a website or place where I can track these papers when they are published, which media channel is best to notify me when these are published?
If it weren't for the missing data, I'd approach this problem from a different angle: since movement is often sudden with lots of "dead space" in between, I would attempt to craft a temporal seam carving algorithm to consolidate the frames where little is happening while preserving the gradual changes. Combine that with optical flow to compensate for things like camera movement, and you should have a much more stable base from which to further refine the output.
While installing this, has anyone else been having an error involving MSBuild.exe? I have the correct 2019 version installed but can't seem to get around it.
also, where does 2 minute papers get his papers?
trees
Pure magic. Are there AI plugins for Photoshop & compatible out there?
It's crazy how far Nvidia is pushing this tech and I'm so grateful that they share these papers with us
Great video. Can you please cover more AUDIO papers?
I love how "Hold on to your papers..." got a pop-up emoji 😂, Great content as always Two Minute Papers 👌
You’re so enthusiastic! Love your videos
what is this program running this where you are modifying inputs?
2 min papers becomes a 7 minute, then I would like to play it on speed 3.5, but Google still haven't AI for this. Amazing paper regardless of speed, though
I love how the chairs are just having a great time dancing
The chairs moving around really caught my attention. Creepy. 🙂
I love your voice and accent. I think they're so adorable. Especially when you say, "and...!".
0:12 my house is in the photo Frankfurt city 😳😳😳😳😳😳😳😳😳😳
ME TOO BRUDER !!!!
@@MeetBuddyAi Main Hauptteil
@@Produciones_Basado ja man
That always feels super special when that happens!
Greetings from Frankfurt, Germany! Love your videos, big thanks :)
🤔 BBC and other documentaries use artificial falsh light to capture plant growth in night. It is consistent. For what could be city timelapse used I dont know. I asume the breakthrough is that it understands the lighting changes, and as NVIDIA pushes its DLSS 3 that produce ugly artifacts in games, when adding in between frames for smoother resolution. But as PC reviewers pointed out, games are smoother with DLSS 3 but has a huge latency, as frames must be computed. So adding more computing for DLSS 3 would mean better quality but even worse latency.
Really love this channel! I'm almost 100% sure the narrator voice in those videos are all AI generated?
To be able to generate images using AI with speech/text is amazing! Even simple words can create the most satisfying looking images especially with the progressing technology on graphics. I'm curious to try create one using Bluewillow, what do you guys think?
Oh wow, that really cool!
The march of progress, baby.
I’m thinking of four-dimensional pictures, where you can watch famous landmarks not just in the 3 dimensions of space, but also in the fourth dimension which is time. And all this would be made from the thousands of tourist photos that already exist.
Would be great to see visual comparison of nn topologies driving the changes.
Can I ask how we can use this?
In the end of 2023 Ai will Rule over movie effect and movie maker..
Make my word guys
Now if only we could do this in real time, cyberpunk level augmented reality would be at hand.
It would be nice to have smooth interpolated playback of the views obtained from the travelling Mars rovers.
"What a time to be alive" exactly!
What a time to be alive :)
He sacrifices THE ROOK! bonus points for those getting the reference.
too smooth. looks fakey fakey
Hi Dr KZF
Kan you please make a video explaining how data is saved inside these great algorithms? For example what kind of data structures are used to by style transfer algorithms, and what members do these data structures have (like arrays of ints? arrays of floating points...) representing what aspect of the style? Thank you.
In his earlier videos he may have already explained what you're curious about. I know he has one on "how AI sees" or something similar to that. It should be full of hallucinogenic images of various things like cats and dogs.
And in a way your question doesn't really make much sense, but that's okay. Neural nets don't have traditional programming data structures. You will hear jargon like latent images or latent data when talking about information stored within a trained AI model. This is the weights of the various node connections and their output biases. From these is where the information is stored. But by virtue of them being a net, there is no 1to1 of any individual values and what information they represent. It is a cumulative effect of multiple values together representing that information. And so any 1 individual value is responsible for multiple representations. Think of it as there are multiple layers of "images" laid out in various configurations, across different regions of the model, and they overlap without issue. As long as no 2 "images" are identical you can store as many of such permutations as the size of the model can allow. There are many issues with this explanation of course, but hopefully this better illustrates how to visualize the "data structure" that can be found in AI models
i didnt know manny khoshbin was a scientist... great success
Personally, I love the chairs moving!
This would be awesome for timelapse of kid to adult photos.
It would be great if the AI could control the camera settings too, so that there isn't any loss of data through over/under exposure over time
Yep, we are in a simulation
Looks like gaming frame generation being used in timelapses. Looks good.
Imagine if AMD could do something for their users with machine learning...
What a time to be alive!
Far from perfect -- it literally hurts my eyes to look at, because everything's got some weird ghosting, like I'm seeing double, and my eyes keep trying to adjust to it.
magyar akcentus :) imádom. Jo a csatorna
I love it, thank you for the effort.
You know the future is eventually going to be great once this Technologies implemented for gaming. Imagine God of War, or Untitled Goose game 🤩😂
Thanks for the vid 😄
How to use this thing?
It makes myself so happy that i'm living through this.
Very cool indeed! But it's pretty wobbly aroud plants and it looks liquidy, guess it wil be fine in two papers down the line ;) Still impresive!
this needs to be in video games
Matrix Awakens already got it
why?
Not sure what to think of this. Technically, its more than fascinating.
every real estate photo of rural areas will be beautiful 😂 doesn't matter the weather!
It’s like curing votes during adjudication. Filling in and making up results based on the operator’s expected outcome. 😅
The day an IA will start to make better reviews of IA progress than this channel, I guess he will not longer said "what time to be alive!"
In fact, the IA would be the one saying that :)
By the time an AI knows its a great time for it to be alive, maybe there won't be anyone left for it to talk to, except robots and other AIs:
"What a time for us to be the only ones left alive!". 🙃
Seeing as AI could keep up with and promote it's own inventions, or follow feeds to see what other AIs are up to, this time may soon come. And rather than crediting teams of researchers, it may well be crediting itself.
@@perfectfutures yeah, these last decades our technology has grew up exponentially, but I guess this would be absolutely nothing compared to what is coming, we need to study over 40 years of specialization and experience to achieve (with lucky) one break thought in one specific field.
One AGI can do it in all fields at once and then it can improve itself, first doubling power every year, then every 3 months, then few days until it knows everything that can be know..
I am here thinking in what I should learn next to improve my coding job..
But then I ask myself, what is the point?
How much time we have left before IA can do everything better than us.
We are the biggest bottle neck for any fast progress. IA developers are already finding all ways to skip that bottleneck.
Thank you!👏🏼👏🏽💯
Old McDonald had a server farm, AI AI oh!
I wander if we can use this AI to “smooth out” your voice so it’s suitable for human consumption like those time lapse videos?
Dude! I could make a fan short of Stephen King's The Mist using ai soon enough!
would be nice to see those improvements in middle frame generation in general
AI is now at the forefront of technology as some like mobile phones are somewhat feeling stagnant nowadays just waiting for the next eureka moment. So far, i've been having a lot of fun using AI image generators like Bluewillow in testing out these AIs. Hoping for the best!
Consequences will never be the same.
and how to install and try this out ? anyone experience so far ? :)
I would like to see a building being constructed time lapse by AI.
5:30 look at the mountain-peak though
Every new paper we must to doubt perception of reality more and more.
I would run the pre- and actual time laps still photos of the 1980 Mt. Saint Helen's eruption.
There have been attempts to tween the still into "footage" augmented with CGI but they're HORRIBLE.
This might do a much more interesting job.
Nvidia, Google, Meta, Microsoft, Open AI, Amazon, Apple.
These companies are on AI race and i love it because the consumer will eventually be the winner!
X
Not really... they are working on AI to make more money off of people, they will keep increasing their subscription models, just to make sure that the consumer owns nothing and has to pay regularly for everything they use, alongside the increasing rent, not to mention inflation, food prices, etc... All these companies are evil and they deserve to fall.
Apple definitely isn’t, they aren’t even trying
@@cyancoyote7366 that’s true. But I think if they are many companies in the market, the prices will be competitive.
Think about the monopoly that IBM had before Microsoft ,Apple , Dell etc were launched.
The computer was very expensive back then.
Competition is generally good for the consumer.
@@MasterKey2004 Apple doesn't have an history of showcasing faulty prototypes.They'll let others try and fail at AI UX before launching anything, just like they did with smartphones and vr headsets.
A cool method with code is better than a cool method without code
The results are very good, actually so good that many will accept results as reality; so, what is being developed to identify AI-produced results from reality?
I wonder if it works on those time-lapses of someone growing a beard. Or a child growing up.
Why would we want fake nature videos?
Nvidia's AI needs to give two minute papers a new voice narrator
Next up: NVIDIA's NEW AI: Realtime Videos Generated Realtime with Realtime Inputs without GPU
hmm... i actually don't love the look of this. the details in detailed areas like leaves/branches/grass look like some false morph effect. smooth areas like clouds and fog look great though
AI image generators like Bluewillow accepts reference images and possibly enhance them. In addition to that, editing tools are starting to accept AIs.
it is interesting but I don't understand why can't we just play every frame on higher frame speed?
That is what is happening at 00:29. There is a lot of flickering
It would be possible to teach a Neural based AI, to deconstruct the mind/thought patterns and reconstruct those patterns as an simulated virtual interactive image. This would be an Approximation. AI should be able to learn to translate our thought processes and it would be able to reconstruct these processes to manifest the reconstruction in a Virtual reality in realtime.
That's a little how some AI aided brain-machine interfaces work: you give the AI training data with both EEG readings and a description of what each person was thinking at the time of the reading and the AI learns what electrical patterns in the brain mean, so later it can literally read someone's mind (provided that the person is wearing a helmet full of electrodes 😂).
There is no visual representation from that yet but probably once the adoption of brain implants such as Neuralink becomes massive, we will see huge advances in the interpretation, visualization and simulation of the human mind's processes.
Maybe soon we will be able to explore each other's minds simply wearing VR lens.
Architects: we are all in ;).
How do they even train this? O_o
That's it... this is too amazing. Sign me up, how do I get the required expertise and resources? I have a 4090, I hope I can use that.
Morphing character transformations (eg human to werewolf) in sci-fi/horror movies and games
Ok, I'm nasty, I know...
"If only an AI exist... that... somehow... could make... your speech... smooth as well" 0:44
Sorry, I couldn't resist to make that joke 😂 and it's not meant to hurt your feelings