Quick technique for the expansion to be perfectly smooth : use an expression on the scale property of the layers and type in "[100,100]*Math.pow(3,-value)" where you replace "value" by a link to a slider control effect, varying from 0 to 100.
I remember seeing this "infinite" zoom effect in Contact (Zemeckis, 1997, film's opening sequence) and in Limitless (2011). I think it was also used in other films, but hey, at those times and without the incredible tools we have now, it was already stunning!
@@SpydersByte Oh yes it is!!! The scenario is great, and it includes political elements and a philosophical (and religious) rational. The infinite zoom effect to Ellie's eye is very nice. I wonder what kind of technology they did use back then ?
@@wuxpoint8311 According to Carin-Anne Strohmaier, the first assistant film editor, the shot was created through three different plates, digitally manipulated in CGI to create the effect: one plate was from the cameraman leading Ellie, the second of Ellie opening the cabinet door (which was a blue screen instead of a miror), and the third of the reflection of Ellie and her dad when the door closes. It's actually really simple technology they used. It's so effective because the timing and lighting is perfect.
I like the "stuttering" effect because it lets you keep track of how big each expansion is (or how many there are). It is very trippy though, I'm glad it wasn't much longer than it was.
@@hunterG60k Honestly, after the results, my eyes were doing the "zoomy" thingie with objects and it felt quite similar to some of the visuals of acid lmao
Dude. I see and understand the smoothness factor you weren't 100% happy with, but syncing it with that music beat made the world's best lemonaid out of the lemons you were given. Nicely done!
Now we know where Joe's secret hideout is... In a cave, in a mountain, at the bottom of the sea, in the middle of the universe, in a rock, in a desert, in a road, in a pebble, in a bay on the side of a cliff, in a lagoon
Came looking for this comment. At 15:05, the beach became a rock, and the sky became the ocean. Joe reinvented the "Honey I Shrunk the Kids" shrinkray, then took a "microscopic" picture with an AI telescope. 😄
I really think that at some point the Earth should have started curving, then we zoom off the planet, out of the solar system, etc., not that weird folded universe thing, AND THEN BACK ON EARTH! Or was it?!
The reason the “Wawawa” effect is happening is because you’re going from zooming out to the edges of an image to zooming out of an image zoomed directly in the center with a constant speed. It’s a similar effect as how moving down a corridor looks slower if you focus on the middle and block out the walls close to you, so as you zoom out the edges of the image get unblocked and the zoom feels faster and faster until the cut to the next image where you’re zoomed in again. The way to fix it would be to start the zoom fast for each image, and then ease it out into a slow zoom. I think that would make it feel more like a continuous motion, but I am gonna have to test it myself at some point to make sure. Very cool experiment and I still love the end result!
_> It’s a similar effect as how moving down a corridor looks slower if you focus on the middle and block out the walls close to you, so as you zoom out the edges of the image get unblocked and the zoom feels faster and faster until the cut to the next image where you’re zoomed in again._ Interestingly, it's very similar to why your brain perceives the moon as larger when it's closer to the skyline.
To be honest, the app "zoom out" looks like a pre-generated effect, with your photo placed in the middle. There is barely any overlap or continuation of the original photo - it immediately went cyberpunk for no reason at all.
What [ColdFusion] describes as "AI washing". Claiming your technology uses state-of-the-art generative AI when in reality its output looks more like it was generated by a content aware fill algorithm. Much like the bricks appearing behind the sheep in a barn, the zoom out app ignored the image context and just grew a texture from what was along the edges of the frame.
@@bornach To be completely fair, this was a very extreme stress test to begin with. Taking a photo that captured a small and relatively complex space and tripling the size of that photo is just not something you would normally do with a feature like that. Secondarily, the setting was also a bit of a niche compared to the more generalised dataset it was trained with. Food, people, animals, plants, and the outdoors... some of the very things that Photoshop users edit the most frequently. As opposed to a studio belonging to a content creator. There are obviously not a lot of images of content creators' studios, and even if there were, they would still be very difficult to replicate since they vary dramatically and can be very personalised. These studios can go all the way from being in the corner of a bedroom to being an entire warehouse. Hell, some even give viewers the illusion of being in a naturally lit and cozy home environment when they're actually just sitting in a small, purpose-built set surrounded by black walls. Sometimes it's not even that, but just a high-quality greenscreen implementation. Even we can fail at predicting how things look like behind the camera of these studios, it's no wonder that deep-learning AI does the same. This is not state-of-the-art AI - it's years behind what companies like OpenAI and Nvidia are currently achieving - but it's still leagues ahead of what content-aware fill algorithms did.
I tried to look closely at the "bricks", and I think it was actually the sheep's curly grey wool being mistaken for part of the wall. Also with the "cyberpunk look", the colorful decorations in the room around the edge of the "original" got stretched into a "80s Futurist" look. It wasn't for "no reason at all", it just used the outer edges instead of a wholistic view of the entire photo. When AI is capable of that, I'll be more impressed.
@@squirlmy I believe Photoshop AI uses the actual photo as a reference. I believe it will recreate images in the main image if you prompt them with a description of the image.
The weird thing about this, and I've tried it, is that after _200_ expansions, you will end up outside a Chuck E Cheese in San Bernardino. Every time. The same Chuck E Cheese. And it's actually there, I found it on the map, at least, last time I checked.
Just like how that consistent image of a woman's face keeps popping up when people use negative prompts. She even has a name that keeps appearing in the text and titles - Loab. Nexpo did a video - "The disturbing art of ai", I think it's called.
17:59 can u not even recognize that it also sucked? like who cares about how far it zoomed out, the first “stage” of zoom out immediately did not blend with your photo in any way. it just generated a weird cyberpunk style building and then put ur picture in the middle of it
It went for a cave because the first shot was clearly an interior space - but theres not many places you can be so far away from someone but still inside except a huge natural cave.
That's a common mistake in games/real-time graphics too. To make the zoom look smooth you don't want to use a linear or cubic curve like a lot of animation packages provide, you need an exponential curve. Zooming out by a factor of 3 like you are doing would mean 3^-x where x goes from 0 to 1 for each segment. Fun fact, 3^100 is about 5x10^47 which is... huge! Like the difference in size between subatomic stuff and the size of the observable universe huge. It's baffling how a smooth, understandable zoom like that compounds so quickly isn't it?
@@waffles_plays5667 Say you want to zoom in so that every second objects in the scene look twice as big. You'd start by zooming from 1x to 2x the first second, then from 2x -> 4x, then to 8, 16, 32, 64, etc. This is exponential, and the numbers get big _very_ fast. (After only 20 seconds the zoom factor would be over a billion!) The curve is zoom=2^time. To animate something like this you really just need the first part of the curve, and then repeat it over and over again for each image. The problem is that it's hard to approximate by hand with animation curves without just calculating the values and throwing them in as keyframes or something.
vertical photos are the bane of any photographer or videographers existence because its not natural. I hate seeing all the appliance users take vertical photos with their phones. Just crap.
I feel like it would’ve made sense to have the pictures be replaced rather than merged. As you zoom out you just remove the smaller picture at the exact moment it lines up with the larger picture.
I was thinking this exact thing. Start zoomed into the middle of the first expanded image. Zoom out until the entire image is in frame and replace it with the next image zoomed in. They must have thought of this, surely. Maybe the images need to be swapped a bit sooner to allow for the overlaps, but still ... ?
I know it probably doesn't matter but the magic is mostly placed in the lens, not the camera itself. So then for this, it would be the best zoom lens. No point in getting technical but do you ever just get that weird feeling over something that doesn't matter yet you feel the need to say it? I'm a strange individual so maybe it's just that. Hope you are having a nice day!
Right after the zoom, just when he says "Is my face all swirly now", stare at something in the room other than the screen and it will swell up bigger and bigger. Joe has effected your brain.
@@NickTaylorRickPowers Uh... notice I say implant it in an elephant? :) There was talk a few years back about taking genetic material from a frozen wooly mammoth, altering an elephant embryo, and implanting in an elephant until birth. This would only be feasible if they can find enough intact DNA. Like Jurassic Park but for a different era. :)
Hey Joe, quick tip to get you from lv7 to lv8 in Ps. Any transformation, crop included, you can press the Alt/Option key while dragging the handle of the transformation to resize it from the central anchor point. Basically you wouldn’t have to drag the corner and re-centre the image over and over. So that’s Alt/Option + dragging a corner of your transform frame
The interesting thing to me was that actual expansion seemed to stop very quickly, replaced by backing off. In other words, instead of zooming out it moved away from the subject in z space. Especially in the one you did in Photoshop, I kept waiting to see the curvature of the Earth, followed by the entire planet, then the rest of the solar system, etc. But nothing like that happened - instead it just showed more landscapes entering the picture at basically the same scale.
thats because its just trying to expand the image, its not trying to zoom out, nor does it have any clue that we're on Earth or that if you zoom out from a landscape like that youll eventually see a curve and then the whole Earth. It has no clue about any of that. It's just an image creation tool, it has no human knowledge whatsoever just a bunch of images to generate from.
The expansion stuttering could be solved by pre composing the composition and time remapping the pre-comp. Would love to have a crack at it when I get the time to! Also, I don't think the AI imaging started with Dall-E, there was an earlier online tool where you could paint an image with basic colors, like using MS Paint, and then it would generate your painting into a "realistic" image. I remember you could do animals, landscape and for some reason.. purses. This was a couple years before Dall-E was revealed, and was most probably one of the earlier cases of AI imaging. Loved the video, Joe!
The first technology that was used for this was Google's deep mind. But it was less image generation and more parameter optimization. Deep mind was all about taking a single image and increasing the amount of x inside the image. Words could be dogs or cats or birds or whatever the classification data was on the original image. That's where you got the crazy photos where they would just up the amount of "dogs" in the image and would give you the trippiest stuff 😂
One of the main reasons that you had difficulties with this is that you're confusing zooming out (Which Photoshop *sort of* manages to do) with trucking back (which it totally can't). With a real camera, if you took a series of images with a successive .5 x zoom like that, with your starting image, you'd need an extreme fisheye device of some sort, and after about half a dozen iterations or so, you'd have reached the limit of a 360 degree panoramic image.
9:25 - Some of those Hand hallucinations are fantastic. 🤣 You owe your friend more than a few beers for composing that zoom sequence. That was awesome.
You know what finally fixed the human anatomy related image generation issues? The "F" Stable Diffusion models _("F" is for "Female")_ since Stable Diffusion is open-source. _(You may look it up in an incognito tab using a public Wi-Fi now. \s)_ Then, the base models began merging some weights from those more _interesting_ custom models/did the same to improve theirs. What I'm telling here applies to all the image generation models, and not only to Stable Diffusion. And yes, another similar thing is taking place with the open-source LLMs also. They're improving wonderfully fast.
Btw, they could've stitched together that big video in the reverse order and then reversed it back for the same effect for way cheaper & quicker. All that it'd have taken them to do so was to start with the (N)th image, place the (N-1)th image in the middle 1/9th section of it, zoom in by 3x (& fade in the full quality (N-1)th image at the end), repeat until the original image. The best part? Even a "dumb" video editor can do this whole thing xd.
I'm into photography and the Generative Fill tool on Photoshop is very useful. While I find the generation of an image via prompt (ala Midjourney or Firefly standalone) to be underwhelming and a novelty, the ability to take my own photos and improve them with the tool is a game changer. Usually it's just dealing with a distracting element or removing tourists from a landscape photo of a rock arch in Utah. But there's also fixing the edges and crops. but like I said, there's nothing fulfilling about generating an image from a text prompt. Sure, I can generate a flawless portrait, but the experience of finding someone to be my model and having an actual human interaction during the shoot process is deeply fulfilling. AI is going to flood the world with disposable, endless images and music and text. Because of that, work created by actual humans will gain scarcity and there's always value in scarcity.
Fun! I’m an editor and dabbling more and more in Photoshop and After Effects. I was chatting with a colleague the other day about an endless zoom out video shared by NATO that was breaking my brain. I wonder if the finger pinch zoom out helps avoid that stutter-step problem your friend had. Lastly - have you ever done a video about early “photoshop”? Like photo manipulation done in the time of early photography? Like those famous photos of Grant and Lincoln at civil war battlefields (but it’s actually their heads were cut out and placed on a person in the shot) or cards displaying “mediums” performing their talents (like ectoplasm coming from nostrils - although maybe that’s a very different video). Good stuff, keep it up Joe!
@@christophstahl8169 Oh man, no layers? Hard to believe that even "undo" was kind of a late concept too. Can you imagine drawing some complex drawing with no layers AND no undo? Yikes.
The Cottingley Fairies would be good to include in something like this, in 1917 two young girls in England took some photographs they claimed showed real fairies, many years later they admitted they were just paper cut-outs. Not exactly photoshopping, just in-camera trickery, but it's one of the most famous cases of it in the UK.
Scaling tip! In After Effects you can parent all your objects to a Null object, and scale only that. Think of it like a train pulling all the railcars that are attached to it, but you only need to animate the train. Essentially you create your null object, it will be auto centered. Parent your first photo to the Null. Now add your second photo on top, set the blend mode to Difference. Scale your Null down to where you see the photos align. Now parent Photo 2 to the Null. Import Photo 3, repeat the process. You’ll have to animate opacity of the overlaying layer from 0 to the 100 during the zoom to make the transition be smooth. Also set your layers back from difference to normal between each new layer setup. Null objects are the best!
could you also layer all the images on top of each other, scaled right, then create a new comp out of that so then its turned all those images into one image, and then zoom out from that?
@@RES1978 I'm such a n00b I actually checked the like count and thought, am I missing something?! And then only it dawned on me... Duh... And that even without being high ;-)
To get rid of the stutter, you could cut the last frames of the expansion of each stitch and then use RIFE interpolation to smooth out the transition. you would get some artifacts, but most likely would work.
Joe's office is his man cave. So putting him in a cave makes sense. That last expansion was cool! I was expecting it to end up showing the entire Earth.
This is an excellent visual example of what is going to happen as AI content becomes more ubiquitous, and new models are trained off that content. AI content feeding AI content....an AI content entropy.
Pretty sure this ended up happening with an AI "art" program, and the results were turning out worse than the previous ones, so basically the AI was inbreeding and it came with all the bad side effects from IRL inbreeding.
Nice! Quick tip: hold shift while expanding your image generative frame from the corner of it. This will make the picture stay in the center of your selection and eliminate for you the need to drag it there manually afterwards
I appreciate this video. As much as this term upsets some people I know, this is one of those "how the sausage is made" videos and I'm glad these get posted. 5:41 Even the lighting is decent.. 6:32 Is that a fridge to the right, or some sort of post-apocalyptic "life-pod"? 7:02 It's looking more and more like some sort of WWIII era self-sustaining bunker.
What if you only filled a 1 pixel wide border around each new image? Would something actually take shape, or would it be incomprehensible? One pixel might be too small, but you could try it again at different increments to see what effect it has.
Honestly, the stutter was great, possibly bc of the chosen music and timing of the beat with the stutter, but I think that was way better than it would’ve been if it was just smooth.
This seems like a replicator from Star Trek but for images instead of food. Saving time and having all the ingredients to make anything you ask it to make.
There is a huge debate in universities right now about how to handle AI because it’s too easy for students not to learn the information being presented in classes and use AI to cheat on assignments instead. My professor was agonizing over the fact that she may need to start doing oral exams with students 1 on 1 in order to combat AI.
I quit university teaching 6 months ago. The last 12 months were a comic slog through a hellscape of AI-produced papers. By the end, I spent more hours per week documenting cheating cases than actually helping the 20% of students who wanted to learn. Of course I’d hoped AI would help me grade student papers but it can’t do anything requiring intelligence! I understand why students use it, but AI sucked all the remaining fun out of academia for me. AI should give us all a chance to re-evaluate modern universities, and the ways they exploit both students and professors, but there’s too much money in those endowments, it will never happen. 😢
@jessamineprice5803 I do not doubt that for a second. I am taking an online course my my Master's in Aviation (my employer is awesome enough to reimburse me for it). It is not hard material, especially for someone who has a bachelor's degree and/or is in the aerospace industry. Despite the ease, I am *positive* some of our required discussion topics are being done by AI (well, student using AI). If the professor is doing something about it, I'm not seeing it. It sucks when you are actually putting effort into something, or you are actually enthusiastic about how airports fund themselves, and you see this canned, boring (sometimes wrong) stuff all over the place.
You did a reasonably good job of covering why firefly is different from the other generative AI platforms, however there’s one thing you didn’t mention, and it’s sort of important. After complaints, Adobe did provide payment for contributors to Adobe stock. It’s unclear if that Money will be repeated, or if it’s just a one time payment. Also, after complaints, Adobe also included the ability to have your art removed from their training material if you want. However… If you remove your works from the training system, you can no longer sell your work on Adobe stock. What this means is that it’s kind of like a sort of extortion. If you want to make money off of one of the most popular stock services in the world, you have to allow AI to scrape your images and create competing work. For a lot of people, that’s going to be a very difficult decision. do you leave your work somewhere where it is the easiest to find, but risk helping this service make you obsolete? Or do you take your work off of the service, making it significantly more difficult for people to find your work? What’s more, I’m fairly certain that once the lawsuits about artist’s work being used in generative AI systems has been resolved, I would imagine that every stock photography site of any size will follow Adobe’s lead. While this is in theory a more ethical way to train a generative AI, the way that Adobe is doing it effectively forces artists to submit their work to make themselves obsolete. A better policy would be to allow artists to opt out but still sell their work. I would even consider limitations on either number of works, repetition of similar works, or some other reasonable limit to be fine so long as the artist could choose to not be included in the training, but still be able to sell on the platform. As of my writing this comment, I’m unaware that Adobe has made such a change. Additionally, it is still unclear if Adobe is going to pay artists more than once. At minimum, if they are going to require works to be used in training, there should be an ongoing payment for either every month or every year that an artist’s work is being used to train firefly.
Also they recently got caught training their models on other AI image generators so they're just as unethical as everything else. They've also started leaning on people pretty heavily to upgrade to the versions with firefly bundled with Photoshop by having screen covering popups spam older versions telling you you need to upgrade. So I'm going to transition to another program in the future. I suspect that one of the reasons that they want people to use the newer versions is so that they can train the AI off of more people's work although I can't substantiate that.
@@damientonkin are you suggesting that if you use Photoshop, Photoshop will be looking at what you’re working on and use that as part of its training without letting you know?
While not moral, this reminds of exactly the same thing in the music biz: if you're starting out and/or struggling, you're going to give up your rights, once established, you're not going to make those kinds of deals. It it kinda "extortion", but honestly I don't know anywhere in creative arts where that doesn't go on. Another field is comic books, including manga. You make bad deals, hopefully you survive long enough to make better deals. 😥
@@squirlmy It's not the same, though. Sure, music artists make bad deals, but most of them are just not getting paid enough, and that doesn't make it right. "Sorry Squirmy, but you signed a lease that means you have to be recorded in your house all the time. You should have read the lease" It's closer to the deals made with rural Black American artists who didn't know what the deal even meant, maybe hadn't heard of recordings before, and they signed away all their rights. It may have been "legal" but that doesn't make it right, and doesn't mean we should just let it happen. But the real problem here is that, unlike those bad deals, or work that was stolen from those artists in the 20s-50s, this can and likely will make it so these folks not only won't get compensated, but very likely won't be able to work anymore in 10 years time. It very likely means that it will be impossible for any artist of any kind to make any money doing art along the way. The only people likely to make any money are those who "hit it big," which will be even harder, and more fleeting as the system just copies them at a pace that won't even give them a day of success. Or, artists will become curiosities, pets like some used to be with the wealthy paying their way and essentially owning them. The age of human art is coming to a close, especially if, like you, we just write it off as "life isn't fair." Well, it's true life's not fair, but it's enough that we can die, get diseases, get in accidents. Do we as people have to make it worse?
well, the wahwah-effect of the zoom-out would be an awesome feature (like a beat visualization) for some trippy music... combined with the trippy video.
TH-cam premium content right here! This is my first time watching you and must i say its very well structured! love the segments everything on point!!! chapters are well structured
This is fascinating, and I'm so glad to hear that Photoshop's generative AI is actually based on the stock library they already owned. Knowing that lets me just sit back and enjoy this.
Im actually sick of AI at this point and the illusion is broken. no joke, 75% of the time, generative language ai is giving you wrong info and presenting it as fact. Gemini at least gives sources. It tried to tell me I was wrong while citing a yahoo blog from 10 years ago at one point... a blog. I can't stress this enough, it's almost always wrong and presents it as fact.
A few things will help with your zoom effect. First is consistency with your GenerativeAI expansion. Instead of manually scaling and re-centering with your crop tool, click on the bottom handle of the crop frame and drag downward while holding shift+option, keeping your old image centered in the frame. That still leaves the issue of scaling inconsistency, but it's better than manual. To fix the scaling issue, you'll have to choose a different method of generative AI since you can't crop at a pre-determined percentage of the original image. 1, select the layer ("make layer" first if it isn't already) and transform the layer to 33.35% (NOT 33.33). 2, ctrl/command click on the layer thumbnail to select the image, then invert the selection. 3, use generative fill and select your desired option. 4, merge your layers into one layer. 5, Save your new image. 6, repeat steps 1-6 (I recommend using a Photoshop Action for this instead of sitting there clicking the same buttons 100 times). The next thing that will help the effect is fixing the effect in After effects. Don't use z-space, but use scaling instead. You can parent all 100 layers to a null object and animate the scale of the null, which would apply the same speed of scale to every image. (look up "earth zoom" or "map zoom" tutorials to achieve this exact effect). With that many layers, though, you'd likely run into image buffer issues pretty quick. Alternatively, you could apply a simple two-keyframe scale animation to one layer, select the keyframes, right click, and choose "exponential zoom." Then copy/paste the keyframes to the rest of the layers. To simplify this, stack all layers on top of one another at the beginning of the comp, paste the keyframes to all of them at the same time, then use the "sequence layers" feature to spread them out over time.
To smooth the expansion, make sure that your seed images are exactly the same size and perfectly centred. Expand your first image by 200%. Start your zoom out, with your overlay next image expanded at 200% ready to CUT (not fade) in over the old when the old hits 100%. That way you might reduce the jittering you experienced in your first attempt.
@@joescott Yup, that was as epic as I thought (other than the portrait orientation! ;) - Is that a limitation of a short, out of interest? (Thank you for taking the time to make that btw :D)
I was a concept artist for a small games/film company. Once Ai came in I was fired. I don't blame Ai, as such. More that my value was seen as worthless by my superiors the moment they saw the potential.
An interesting alternative to what you did would have been to write a script to shrink your image by 2px*2px leaving a 1px border around the canvas, then generatively filling that, saving each image as you go. You wouldn't want to do it by hand, but if you left a PC to do this for a few days you'd end up with a crazy animation I think! Increase throughput by increasing margin size, but I think this might sacrifice some quality as it tries to fill the margin with objects from its library rather than creating a 'best fit' pixel giving what's already there.
19:17 "Is it just training off of itself at this point?" Whenever it gives you three options to choose from, I suspect the choice is fed back for further training And so might be the usage (versus lack thereof) of the generated stock images
I respect your standpoint towards the topic 👏(from my experience, most of the guys in videos about using AI are this unempathical type of techbro who thinks everything is better if it's done by a computer... it's nice to have somebody acknoledge the harm caused by AI while still making a point about some benefits)
I like the perspective change on the little rock outcropping in the waves at 15:06 all previous images were on that little rock... imagine how small Joe is inside the cave... like a grain of sand.
Yes you are correct, the way AI is trained is not what many people without knowledge about ML and AI think. It's more like using the data temporarily and then you won't need it again. And it's not like storing the images you train into the model. Also the way it is trained (diffusion model) makes it transformative which makes it not violate copyright law. Stealing image in world is computer is weird. When you think about it, when you steal someone's image, you are really just copying the ones and zeros that they made.
As an artist who has used all traditional media, especially airbrush for over 3 decades, including Photoshop since the early 90's, I have no intention of using Ai features and have avoided even checking them out. I work on my pieces down to the pixel with brush and pen tools as though I'm using what I call analog techniques which is how I differentiate traditional methods from the digital. To me it's just a cheat, as for me "Artists that can, make art while those that cannot, use Ai".
This is just plain old regressive thinking. The same kind of thinking that complained about the printing press, and telephones, and TV, and computers, and internet... Try and be a bit more forward thinking. For a long time photography itself was considered cheating when compared to painting, and then photoshop was considered cheating by photographers. In reality, they're all just tools, and so is AI.
Next time you wanna do stuff like the edit, I recommend using a 3d program, like blender. There, you can just scale the elements based on math, move smaller images up a bit and zoom in using a linear curve. If you truly made the images 300% bigger, you can just scale each image plane with the formula 1/3. Importing the image planes requires an inbuilt but by default disabled addon, "images as planes" Another easier way to do this, would be using stable diffusion and the deforum addon, which has a zoom function. it will alter the original frame, but it is basically how all those trippy zoom videos were made.
I can’t honestly say that any of this stuff shows any signs of intelligence, artificial or otherwise. The photo edges are expanded using an algorithm that essentially copies what is already in the image. Later, larger, expansions are simply matching the image content as closely as possible to another image in a stock library of images. Given that the majority of images in photo libraries are landscapes of one sort or another, it is to be expected that all images will eventually expand into a landscape. So what we have is a complex algorithm and a huge library of images but certainly no intelligence necessary to produce these results.
In other words, exactly what I would do, given the time, finesse, patience, and catalogue of images. Robots with AI are going to build an equivalent to the Great Wall of China at some point. Then people like yourself will argue that building structures isn't a sign of intelligence. Do note that there's also a vast catalogue of city images, and the AI did not flip over to it at any point in the zoom.
Initial thoughts: If you are getting a slight distortion on the edge of the images, make the images oversized and crop them. You are scaling the images to 1/3rd during the generation. If you scale them back 3x in Premiere/AE and then overlay the original image in the middle, you should avoid the non-linear scaling issue. You could nudge the generative fill a little by pushing the image a few pixels off centre and telling it to rotate up etc You could potentially get a zoom out to space if you wanted. Edit to add... motion blur would make the zoom better imho once you fix the non-linear zoom issue.
There's a way easier method to pull this off in After Effects without getting into 3D space. Stack all your images, create a null control layer (start with the scale at something like 10,000) and parent the first image to it, scale down (on the null layer) until the first image fits within the second image (you may need to shift the position of this second image to line them up) and once they're lined up parent the second image to the null layer. Repeat on and on. And when you run out of scale margin (you will be scaling from 10,000% down to 1%) just create another null layer to control the first.
There is a simpler way to do the editing on After Effects. But you'd have to work from the last pictures to the first before reversing the shot. Add your last picture to the composition, put a keyframe on scale on the first frame, then move to, let's say 2 seconds later, and scale your picture to 300% (it will add a keyframe). Then add your second picture to the composition and just parent it to the previous one. move to 2 seconds later, the scale of your first picture/layer must now be 900% (if it's a pain in the *ss to calculate each time, you can just take the scale as it is and add *3 just after, it will automatically calculate it), you add the third picture and parent it either to the second or the first (the result will be the same), then move to 6 seconds, scale your first picture to 2700%, add the fourth picture, parent it to any of the other pictures, move to 8 seconds, scale your first picture to 8100% and so on... You will then have a perfectly constant speed zoom in your creation. It might be heavy on ram since, in the end, you will have 100 layers, but it should work thought it can be optimized. After you've done all of this, if you want a smoothest animation, you can just delete all the keyframes that are located between the first one and the last one, then select the two remaining ones and press f9 to activate smooth interpolation between keyframes, you could even just reverse the two keyframes at this point so you would zoom out instead of zooming in, just make sure to don't do that before you have added all your pictures. Also, in premiere, while the keyframe interpolation is on smooth by default, you can have them be set to linear, but i'm not sure how it's done though, maybe by pressing ctrl while you click on the keyframe symbol... or maybe by pressing f9... i'm not as familiar with premiere as i am with after effects. Sorry for my clunky english, i'm french. Hope it can help you with future projects.
The point about the Ai training off of itself is something my old design mentor mentioned to me recently. He had been using it as a shortcut for lower paying clients or people with lower standards, and he had realized that it had been spitting out worse and worse quality. He also that more and more prompts were being censored, which would put restrictions on what it could train off of or generate affecting the quality. However it shouldn't affect it as much as the quality had been showing, meaning it probably has been training off of itself, creating a negative feedback loop.
Everytime I've seen this expanding image effect, it had this wobble between transitions. In any case, your friend did a great job with this weird AI journey.
with the editing issue; you could've also just placed all the images in their right places, and anchor them to each other. scaling the bigger one would've made the smaller one follow right along.
JEZUZ CRUST!!! After watching the edited version my room was literally distorting before my very eyes when I took my eyes off the screen!! Talk about trippy!
Quick technique for the expansion to be perfectly smooth : use an expression on the scale property of the layers and type in "[100,100]*Math.pow(3,-value)" where you replace "value" by a link to a slider control effect, varying from 0 to 100.
I second this
Part of the problem is he positioned each expansion by hand, so they weren't perfectly centered.
@@dranorterAnd perhaps not scaled by a factor of 3 every time as well, to add insult to injury...
@@dranorter You could use a action or write a scripts to automate it, for perfect alignment.
as soon as I saw him cropping by hand I knew it would end up messy
I remember seeing this "infinite" zoom effect in Contact (Zemeckis, 1997, film's opening sequence) and in Limitless (2011). I think it was also used in other films, but hey, at those times and without the incredible tools we have now, it was already stunning!
Contact is such a good movie
@@SpydersByte Oh yes it is!!! The scenario is great, and it includes political elements and a philosophical (and religious) rational. The infinite zoom effect to Ellie's eye is very nice. I wonder what kind of technology they did use back then ?
@@wuxpoint8311 According to Carin-Anne Strohmaier, the first assistant film editor, the shot was created through three different plates, digitally manipulated in CGI to create the effect: one plate was from the cameraman leading Ellie, the second of Ellie opening the cabinet door (which was a blue screen instead of a miror), and the third of the reflection of Ellie and her dad when the door closes.
It's actually really simple technology they used. It's so effective because the timing and lighting is perfect.
@@RandomSime Woah! Many thanks for the info!
It's also in royksopp's eple music video, with collage influence
I like the "stuttering" effect because it lets you keep track of how big each expansion is (or how many there are). It is very trippy though, I'm glad it wasn't much longer than it was.
Sync it to the beats of the music. But the way he tweaked each one by hand is ... the opposite of saving time.
No way! I was thinking I need 10 hours of this next time I'm dropping acid lol Loved it
You see the stuttering on DMT distinctly before you enter the dome
@@hunterG60k Honestly, after the results, my eyes were doing the "zoomy" thingie with objects and it felt quite similar to some of the visuals of acid lmao
@@slovakthrowback3738I have HPPD and it super tripped me out 😵💫
15:05 is an insane transition, it looks like the entire thing was a miniature world ontop of a few loose rocks
The sky suddenly became the water
how much have you smoked??
Not just loose rocks but the grit of a sheet of corrugated metal.
"Is my face all swirly now?
👁️👄👁️"
🤣🤣🤣 Yes, yes it is.
LMAOOO
👁️
....👄...👁️
I had to re watch that section after the end to make sure it wasn't an editing effect.
😂 is cringe now. Use 💀 or 😭
@@meoutpeaceErm actually, they used “🤣”
Dude. I see and understand the smoothness factor you weren't 100% happy with, but syncing it with that music beat made the world's best lemonaid out of the lemons you were given. Nicely done!
Yea it went well with the beat if the music, more hypnotic and freaky
Now we know where Joe's secret hideout is... In a cave, in a mountain, at the bottom of the sea, in the middle of the universe, in a rock, in a desert, in a road, in a pebble, in a bay on the side of a cliff, in a lagoon
It's Everything, Everywhere, All At Once!
That starts to sound like a Dr. Seuss poem.
Came looking for this comment.
At 15:05, the beach became a rock, and the sky became the ocean. Joe reinvented the "Honey I Shrunk the Kids" shrinkray, then took a "microscopic" picture with an AI telescope. 😄
We should send a postcard to him with that for an address and see if the postal workers manage to deliver it.
I really think that at some point the Earth should have started curving, then we zoom off the planet, out of the solar system, etc., not that weird folded universe thing, AND THEN BACK ON EARTH!
Or was it?!
I don’t need photoshop, I expand myself with McDonalds
Lmao
XD
At the speed prices are going up at the Micky Dees you are better off with the 659.88/yr subscription to adobe.
LOL
GEIJUTSU WA BAKAHATSU DA
The reason the “Wawawa” effect is happening is because you’re going from zooming out to the edges of an image to zooming out of an image zoomed directly in the center with a constant speed.
It’s a similar effect as how moving down a corridor looks slower if you focus on the middle and block out the walls close to you, so as you zoom out the edges of the image get unblocked and the zoom feels faster and faster until the cut to the next image where you’re zoomed in again.
The way to fix it would be to start the zoom fast for each image, and then ease it out into a slow zoom. I think that would make it feel more like a continuous motion, but I am gonna have to test it myself at some point to make sure.
Very cool experiment and I still love the end result!
_> It’s a similar effect as how moving down a corridor looks slower if you focus on the middle and block out the walls close to you, so as you zoom out the edges of the image get unblocked and the zoom feels faster and faster until the cut to the next image where you’re zoomed in again._
Interestingly, it's very similar to why your brain perceives the moon as larger when it's closer to the skyline.
Thanks!
To be honest, the app "zoom out" looks like a pre-generated effect, with your photo placed in the middle. There is barely any overlap or continuation of the original photo - it immediately went cyberpunk for no reason at all.
What [ColdFusion] describes as "AI washing". Claiming your technology uses state-of-the-art generative AI when in reality its output looks more like it was generated by a content aware fill algorithm. Much like the bricks appearing behind the sheep in a barn, the zoom out app ignored the image context and just grew a texture from what was along the edges of the frame.
@@bornach To be completely fair, this was a very extreme stress test to begin with. Taking a photo that captured a small and relatively complex space and tripling the size of that photo is just not something you would normally do with a feature like that. Secondarily, the setting was also a bit of a niche compared to the more generalised dataset it was trained with. Food, people, animals, plants, and the outdoors... some of the very things that Photoshop users edit the most frequently. As opposed to a studio belonging to a content creator. There are obviously not a lot of images of content creators' studios, and even if there were, they would still be very difficult to replicate since they vary dramatically and can be very personalised. These studios can go all the way from being in the corner of a bedroom to being an entire warehouse. Hell, some even give viewers the illusion of being in a naturally lit and cozy home environment when they're actually just sitting in a small, purpose-built set surrounded by black walls. Sometimes it's not even that, but just a high-quality greenscreen implementation. Even we can fail at predicting how things look like behind the camera of these studios, it's no wonder that deep-learning AI does the same.
This is not state-of-the-art AI - it's years behind what companies like OpenAI and Nvidia are currently achieving - but it's still leagues ahead of what content-aware fill algorithms did.
I tried to look closely at the "bricks", and I think it was actually the sheep's curly grey wool being mistaken for part of the wall. Also with the "cyberpunk look", the colorful decorations in the room around the edge of the "original" got stretched into a "80s Futurist" look. It wasn't for "no reason at all", it just used the outer edges instead of a wholistic view of the entire photo. When AI is capable of that, I'll be more impressed.
@@squirlmy I believe Photoshop AI uses the actual photo as a reference. I believe it will recreate images in the main image if you prompt them with a description of the image.
i think it's cuz the ai is too primitive
The weird thing about this, and I've tried it, is that after _200_ expansions, you will end up outside a Chuck E Cheese in San Bernardino. Every time. The same Chuck E Cheese. And it's actually there, I found it on the map, at least, last time I checked.
Just like how that consistent image of a woman's face keeps popping up when people use negative prompts. She even has a name that keeps appearing in the text and titles - Loab. Nexpo did a video - "The disturbing art of ai", I think it's called.
@@jacoboneill2494oh my god my friend was one of the ones who discovered Loab!
After 200 expansions, you end up with a Tool album cover
So, that Chuck E Cheese must be the Alpha and Omega point of the Universe. I feel a cult starting.
This sounds like a SCP concept
Towards the end, I was reminded of a Mandelbrot Fractal. Glad that you are doing something JUST FOR FUN!! ❤
17:59 can u not even recognize that it also sucked? like who cares about how far it zoomed out, the first “stage” of zoom out immediately did not blend with your photo in any way. it just generated a weird cyberpunk style building and then put ur picture in the middle of it
It went for a cave because the first shot was clearly an interior space - but theres not many places you can be so far away from someone but still inside except a huge natural cave.
That's a common mistake in games/real-time graphics too. To make the zoom look smooth you don't want to use a linear or cubic curve like a lot of animation packages provide, you need an exponential curve. Zooming out by a factor of 3 like you are doing would mean 3^-x where x goes from 0 to 1 for each segment. Fun fact, 3^100 is about 5x10^47 which is... huge! Like the difference in size between subatomic stuff and the size of the observable universe huge. It's baffling how a smooth, understandable zoom like that compounds so quickly isn't it?
Can you please explain this to me like I'm a kindergartner?
@@waffles_plays5667 Say you want to zoom in so that every second objects in the scene look twice as big. You'd start by zooming from 1x to 2x the first second, then from 2x -> 4x, then to 8, 16, 32, 64, etc. This is exponential, and the numbers get big _very_ fast. (After only 20 seconds the zoom factor would be over a billion!) The curve is zoom=2^time. To animate something like this you really just need the first part of the curve, and then repeat it over and over again for each image. The problem is that it's hard to approximate by hand with animation curves without just calculating the values and throwing them in as keyframes or something.
@@slembcke Thank you for coming out after 3 months to reply to me :) I sort of understand now.
Thank you kind internet stranger.
@@waffles_plays5667this is actually wholesome ngl
Are you telling me one man built this entire youtube channel from scraps in a CAVE?
He got most of his parts from that mammoth you saw at the beginning. That's actually Joe in the photo, eying up the beast for materials.
Mr. Stane, I'm going to need you to calm down.
The truth is... I am Joe Scott.
Lmfao
vertical photos are the bane of any photographer or videographers existence because its not natural. I hate seeing all the appliance users take vertical photos with their phones. Just crap.
Good lord I’ve never made it to a video 37 seconds after it came out the oven. I’m gonna need the oven gloves.
Hehe, it came out late today so you got lucky!
@@joescott sounds about right, of course I’m only early when the video is late 😂
@@joescottseriously joe I have been asking the name for this 19:47 music quite some time could you at least give the name?
The "Ove" Glove
Loved the Eclipse live stream.
I feel like it would’ve made sense to have the pictures be replaced rather than merged. As you zoom out you just remove the smaller picture at the exact moment it lines up with the larger picture.
I was thinking this exact thing. Start zoomed into the middle of the first expanded image. Zoom out until the entire image is in frame and replace it with the next image zoomed in. They must have thought of this, surely. Maybe the images need to be swapped a bit sooner to allow for the overlaps, but still ... ?
Maybe due to bad quality since it’s lower resolution
14:31 - woah, dude! That was genuinely beautiful. Great track pick too.
14:30 What flat-earthers think will happen if you gave them the best Nikon camera.
I know it probably doesn't matter but the magic is mostly placed in the lens, not the camera itself. So then for this, it would be the best zoom lens.
No point in getting technical but do you ever just get that weird feeling over something that doesn't matter yet you feel the need to say it? I'm a strange individual so maybe it's just that. Hope you are having a nice day!
lmao.
@@wingedfalcon 😂
!!!!!!!!!!!!! TOO FUNNY, but also totally true.
That's what I was thinking. This is what it would look like if the Earth were flat and you just kept backing up.
Right after the zoom, just when he says "Is my face all swirly now", stare at something in the room other than the screen and it will swell up bigger and bigger. Joe has effected your brain.
wow, was not at all expecting that to work but it did! Very trippy!
reminded me of visuals i had when i was younger on acid lmao
When it faded to black, I saw my reflection in the screen and noticed it happening. Kind of freaked me out because I was not expecting it.
🤯 what do you mean that's not an actual photo of a wooly mammoth !?!?! 😂😂😂😂
Any day now they are going to grow one in a test tube and implant in an elephant for birth. :) Then He can get his actual photo.
Unfortunately they weren't able to train the AI on any actual photos of wooly mammoths due to copyright reasons.
@@brianh9358how big is that test tube gonna be
FAKE NEWS...!!!
@@NickTaylorRickPowers Uh... notice I say implant it in an elephant? :) There was talk a few years back about taking genetic material from a frozen wooly mammoth, altering an elephant embryo, and implanting in an elephant until birth. This would only be feasible if they can find enough intact DNA. Like Jurassic Park but for a different era. :)
This is my favorite video of yours! Seeing it turn your place into a cave and exit into a landscape was wild
Hey Joe, quick tip to get you from lv7 to lv8 in Ps. Any transformation, crop included, you can press the Alt/Option key while dragging the handle of the transformation to resize it from the central anchor point. Basically you wouldn’t have to drag the corner and re-centre the image over and over.
So that’s Alt/Option + dragging a corner of your transform frame
Man this legitimately messed with my eyes. I looked down at my dinner right after and it freaked me out a bit.
The interesting thing to me was that actual expansion seemed to stop very quickly, replaced by backing off. In other words, instead of zooming out it moved away from the subject in z space. Especially in the one you did in Photoshop, I kept waiting to see the curvature of the Earth, followed by the entire planet, then the rest of the solar system, etc. But nothing like that happened - instead it just showed more landscapes entering the picture at basically the same scale.
thats because its just trying to expand the image, its not trying to zoom out, nor does it have any clue that we're on Earth or that if you zoom out from a landscape like that youll eventually see a curve and then the whole Earth. It has no clue about any of that. It's just an image creation tool, it has no human knowledge whatsoever just a bunch of images to generate from.
The expansion stuttering could be solved by pre composing the composition and time remapping the pre-comp. Would love to have a crack at it when I get the time to!
Also, I don't think the AI imaging started with Dall-E, there was an earlier online tool where you could paint an image with basic colors, like using MS Paint, and then it would generate your painting into a "realistic" image. I remember you could do animals, landscape and for some reason.. purses. This was a couple years before Dall-E was revealed, and was most probably one of the earlier cases of AI imaging.
Loved the video, Joe!
looks like it was Nvidia Canvas
The first technology that was used for this was Google's deep mind. But it was less image generation and more parameter optimization. Deep mind was all about taking a single image and increasing the amount of x inside the image. Words could be dogs or cats or birds or whatever the classification data was on the original image. That's where you got the crazy photos where they would just up the amount of "dogs" in the image and would give you the trippiest stuff 😂
That was a very interesting idea and video, cheers to you and Mark for putting it together for us
One of the main reasons that you had difficulties with this is that you're confusing zooming out (Which Photoshop *sort of* manages to do) with trucking back (which it totally can't).
With a real camera, if you took a series of images with a successive .5 x zoom like that, with your starting image, you'd need an extreme fisheye device of some sort, and after about half a dozen iterations or so, you'd have reached the limit of a 360 degree panoramic image.
9:25 - Some of those Hand hallucinations are fantastic. 🤣
You owe your friend more than a few beers for composing that zoom sequence. That was awesome.
You know what finally fixed the human anatomy related image generation issues? The "F" Stable Diffusion models _("F" is for "Female")_ since Stable Diffusion is open-source.
_(You may look it up in an incognito tab using a public Wi-Fi now. \s)_
Then, the base models began merging some weights from those more _interesting_ custom models/did the same to improve theirs. What I'm telling here applies to all the image generation models, and not only to Stable Diffusion.
And yes, another similar thing is taking place with the open-source LLMs also. They're improving wonderfully fast.
Btw, they could've stitched together that big video in the reverse order and then reversed it back for the same effect for way cheaper & quicker. All that it'd have taken them to do so was to start with the (N)th image, place the (N-1)th image in the middle 1/9th section of it, zoom in by 3x (& fade in the full quality (N-1)th image at the end), repeat until the original image. The best part? Even a "dumb" video editor can do this whole thing xd.
JESUS IS COMING VERY SOON🔥 REPENT AND TURN AWAY FROM YOUR SINS 🙏🙏🙏😢😢
I'm into photography and the Generative Fill tool on Photoshop is very useful. While I find the generation of an image via prompt (ala Midjourney or Firefly standalone) to be underwhelming and a novelty, the ability to take my own photos and improve them with the tool is a game changer. Usually it's just dealing with a distracting element or removing tourists from a landscape photo of a rock arch in Utah. But there's also fixing the edges and crops.
but like I said, there's nothing fulfilling about generating an image from a text prompt. Sure, I can generate a flawless portrait, but the experience of finding someone to be my model and having an actual human interaction during the shoot process is deeply fulfilling.
AI is going to flood the world with disposable, endless images and music and text. Because of that, work created by actual humans will gain scarcity and there's always value in scarcity.
Fun! I’m an editor and dabbling more and more in Photoshop and After Effects. I was chatting with a colleague the other day about an endless zoom out video shared by NATO that was breaking my brain. I wonder if the finger pinch zoom out helps avoid that stutter-step problem your friend had.
Lastly - have you ever done a video about early “photoshop”? Like photo manipulation done in the time of early photography? Like those famous photos of Grant and Lincoln at civil war battlefields (but it’s actually their heads were cut out and placed on a person in the shot) or cards displaying “mediums” performing their talents (like ectoplasm coming from nostrils - although maybe that’s a very different video).
Good stuff, keep it up Joe!
I seem to remember that Photoshop had its origins in the special effects for "The Abyss".
When I started using Photoshop we did not have layers... so even a video about actual "early Photoshop" could be fun :D
@@christophstahl8169 Oh man, no layers? Hard to believe that even "undo" was kind of a late concept too. Can you imagine drawing some complex drawing with no layers AND no undo? Yikes.
The Cottingley Fairies would be good to include in something like this, in 1917 two young girls in England took some photographs they claimed showed real fairies, many years later they admitted they were just paper cut-outs. Not exactly photoshopping, just in-camera trickery, but it's one of the most famous cases of it in the UK.
That was beautiful, i like the timing of the zoom stutter with the music, came together so wonderfully!
I’ve seen videos just like that sequence for so long! I love them! It’s mesmerizing!
Always learn something intriguing when I watch Mr. Joe Scott. Keep up the great work/content! Great channel!
Scaling tip! In After Effects you can parent all your objects to a Null object, and scale only that. Think of it like a train pulling all the railcars that are attached to it, but you only need to animate the train.
Essentially you create your null object, it will be auto centered. Parent your first photo to the Null. Now add your second photo on top, set the blend mode to Difference. Scale your Null down to where you see the photos align. Now parent Photo 2 to the Null. Import Photo 3, repeat the process.
You’ll have to animate opacity of the overlaying layer from 0 to the 100 during the zoom to make the transition be smooth.
Also set your layers back from difference to normal between each new layer setup.
Null objects are the best!
Boring.
Exactly what I wanted to say 😂
Or just write an expression and copy paste it on all layers, should be done in half an hour 😅
This person game designs.
could you also layer all the images on top of each other, scaled right, then create a new comp out of that so then its turned all those images into one image, and then zoom out from that?
Far out man i was not expecting that swirly face thing haha... i was like dude did you know i was not 100% sober right now? Haha
hahaha same
I knew because I almost fell down after I watched it the first time. :)
@@joescott with mind altering plants 🪴 come mind altering pictures within pictures zoomy zoom zooms... spiral circle face!
This is the highest comment I think I’ve ever seen,
@@RES1978 I'm such a n00b I actually checked the like count and thought, am I missing something?! And then only it dawned on me... Duh... And that even without being high ;-)
To get rid of the stutter, you could cut the last frames of the expansion of each stitch and then use RIFE interpolation to smooth out the transition. you would get some artifacts, but most likely would work.
Watching the final cut of it all and staring at the center makes things morph when you look away. So trippy
For AI to put you in a cave, as if you lived in a cave or under a rock was beautiful. It knows us humans too well
Joe made this video in a cave with a box of parts!
Joe's office is his man cave. So putting him in a cave makes sense. That last expansion was cool! I was expecting it to end up showing the entire Earth.
This is an excellent visual example of what is going to happen as AI content becomes more ubiquitous, and new models are trained off that content. AI content feeding AI content....an AI content entropy.
Pretty sure this ended up happening with an AI "art" program, and the results were turning out worse than the previous ones, so basically the AI was inbreeding and it came with all the bad side effects from IRL inbreeding.
This is not how AI works are you guys are serious?
AI use CLOSED databases, not the open internet to learn.
@@Kongongonggnot all of them, and over time those "closed databases" will need to be expanded and likely contaminated.
@@logandarnell8946 nah, don’t worry AI pictures will not make it into these dataset for very long time.
"is my face all swirly now?" jumpscared me, because, yes, yes you were swirly
Nice! Quick tip: hold shift while expanding your image generative frame from the corner of it. This will make the picture stay in the center of your selection and eliminate for you the need to drag it there manually afterwards
I appreciate this video. As much as this term upsets some people I know, this is one of those "how the sausage is made" videos and I'm glad these get posted.
5:41 Even the lighting is decent..
6:32 Is that a fridge to the right, or some sort of post-apocalyptic "life-pod"?
7:02 It's looking more and more like some sort of WWIII era self-sustaining bunker.
What if you only filled a 1 pixel wide border around each new image? Would something actually take shape, or would it be incomprehensible?
One pixel might be too small, but you could try it again at different increments to see what effect it has.
The stagger/stutter actually works well when synced up to music! imo.
Honestly, the stutter was great, possibly bc of the chosen music and timing of the beat with the stutter, but I think that was way better than it would’ve been if it was just smooth.
The ideal case: you end up recreating the 'Royksopp - Eple' music video
This seems like a replicator from Star Trek but for images instead of food. Saving time and having all the ingredients to make anything you ask it to make.
There is a huge debate in universities right now about how to handle AI because it’s too easy for students not to learn the information being presented in classes and use AI to cheat on assignments instead. My professor was agonizing over the fact that she may need to start doing oral exams with students 1 on 1 in order to combat AI.
The universities are finished and there will be no jobs for most people at university today thanks to AI anyway
yea I could definitely see things going in that direction, at least its one of the potential solutions
I quit university teaching 6 months ago. The last 12 months were a comic slog through a hellscape of AI-produced papers. By the end, I spent more hours per week documenting cheating cases than actually helping the 20% of students who wanted to learn. Of course I’d hoped AI would help me grade student papers but it can’t do anything requiring intelligence! I understand why students use it, but AI sucked all the remaining fun out of academia for me. AI should give us all a chance to re-evaluate modern universities, and the ways they exploit both students and professors, but there’s too much money in those endowments, it will never happen. 😢
You should have used a better AI. 😏
@jessamineprice5803 I do not doubt that for a second. I am taking an online course my my Master's in Aviation (my employer is awesome enough to reimburse me for it). It is not hard material, especially for someone who has a bachelor's degree and/or is in the aerospace industry. Despite the ease, I am *positive* some of our required discussion topics are being done by AI (well, student using AI). If the professor is doing something about it, I'm not seeing it. It sucks when you are actually putting effort into something, or you are actually enthusiastic about how airports fund themselves, and you see this canned, boring (sometimes wrong) stuff all over the place.
You did a reasonably good job of covering why firefly is different from the other generative AI platforms, however there’s one thing you didn’t mention, and it’s sort of important.
After complaints, Adobe did provide payment for contributors to Adobe stock. It’s unclear if that Money will be repeated, or if it’s just a one time payment. Also, after complaints, Adobe also included the ability to have your art removed from their training material if you want. However…
If you remove your works from the training system, you can no longer sell your work on Adobe stock.
What this means is that it’s kind of like a sort of extortion. If you want to make money off of one of the most popular stock services in the world, you have to allow AI to scrape your images and create competing work. For a lot of people, that’s going to be a very difficult decision. do you leave your work somewhere where it is the easiest to find, but risk helping this service make you obsolete? Or do you take your work off of the service, making it significantly more difficult for people to find your work? What’s more, I’m fairly certain that once the lawsuits about artist’s work being used in generative AI systems has been resolved, I would imagine that every stock photography site of any size will follow Adobe’s lead.
While this is in theory a more ethical way to train a generative AI, the way that Adobe is doing it effectively forces artists to submit their work to make themselves obsolete. A better policy would be to allow artists to opt out but still sell their work. I would even consider limitations on either number of works, repetition of similar works, or some other reasonable limit to be fine so long as the artist could choose to not be included in the training, but still be able to sell on the platform.
As of my writing this comment, I’m unaware that Adobe has made such a change.
Additionally, it is still unclear if Adobe is going to pay artists more than once. At minimum, if they are going to require works to be used in training, there should be an ongoing payment for either every month or every year that an artist’s work is being used to train firefly.
Also they recently got caught training their models on other AI image generators so they're just as unethical as everything else. They've also started leaning on people pretty heavily to upgrade to the versions with firefly bundled with Photoshop by having screen covering popups spam older versions telling you you need to upgrade. So I'm going to transition to another program in the future. I suspect that one of the reasons that they want people to use the newer versions is so that they can train the AI off of more people's work although I can't substantiate that.
@@damientonkin are you suggesting that if you use Photoshop, Photoshop will be looking at what you’re working on and use that as part of its training without letting you know?
I appreciate that clarification
While not moral, this reminds of exactly the same thing in the music biz: if you're starting out and/or struggling, you're going to give up your rights, once established, you're not going to make those kinds of deals. It it kinda "extortion", but honestly I don't know anywhere in creative arts where that doesn't go on. Another field is comic books, including manga. You make bad deals, hopefully you survive long enough to make better deals. 😥
@@squirlmy It's not the same, though. Sure, music artists make bad deals, but most of them are just not getting paid enough, and that doesn't make it right.
"Sorry Squirmy, but you signed a lease that means you have to be recorded in your house all the time. You should have read the lease"
It's closer to the deals made with rural Black American artists who didn't know what the deal even meant, maybe hadn't heard of recordings before, and they signed away all their rights. It may have been "legal" but that doesn't make it right, and doesn't mean we should just let it happen.
But the real problem here is that, unlike those bad deals, or work that was stolen from those artists in the 20s-50s, this can and likely will make it so these folks not only won't get compensated, but very likely won't be able to work anymore in 10 years time.
It very likely means that it will be impossible for any artist of any kind to make any money doing art along the way. The only people likely to make any money are those who "hit it big," which will be even harder, and more fleeting as the system just copies them at a pace that won't even give them a day of success.
Or, artists will become curiosities, pets like some used to be with the wealthy paying their way and essentially owning them.
The age of human art is coming to a close, especially if, like you, we just write it off as "life isn't fair."
Well, it's true life's not fair, but it's enough that we can die, get diseases, get in accidents. Do we as people have to make it worse?
Love this video. I've been using Photoshop professionally since the 90s and can relate.
well, the wahwah-effect of the zoom-out would be an awesome feature (like a beat visualization) for some trippy music... combined with the trippy video.
That phone app thinks your studio is in Night City.
I like how it turned into barn tin for a minute lol
Now Joe is 1/1000 the size of his mom.
Or 0.001
Or 10‰
@@Schöenebuddy10🐛
🤯🤯🤯
Is that a 'yo mama so fat' joke?
TH-cam premium content right here! This is my first time watching you and must i say its very well structured! love the segments everything on point!!! chapters are well structured
This is fascinating, and I'm so glad to hear that Photoshop's generative AI is actually based on the stock library they already owned. Knowing that lets me just sit back and enjoy this.
Im actually sick of AI at this point and the illusion is broken. no joke, 75% of the time, generative language ai is giving you wrong info and presenting it as fact. Gemini at least gives sources. It tried to tell me I was wrong while citing a yahoo blog from 10 years ago at one point... a blog. I can't stress this enough, it's almost always wrong and presents it as fact.
So, just like the average keyboard warrior on Facebook then?
Imagine the school text books in 2050 that will all be A.I. generated 🙄
As a life-long adobe addict and tech enthusiast, this video is like catnip
And yes modern AI tools are terrifying
Goddaaamn! This was trippy. Very, very trippy.
A few things will help with your zoom effect. First is consistency with your GenerativeAI expansion. Instead of manually scaling and re-centering with your crop tool, click on the bottom handle of the crop frame and drag downward while holding shift+option, keeping your old image centered in the frame. That still leaves the issue of scaling inconsistency, but it's better than manual.
To fix the scaling issue, you'll have to choose a different method of generative AI since you can't crop at a pre-determined percentage of the original image.
1, select the layer ("make layer" first if it isn't already) and transform the layer to 33.35% (NOT 33.33).
2, ctrl/command click on the layer thumbnail to select the image, then invert the selection.
3, use generative fill and select your desired option.
4, merge your layers into one layer.
5, Save your new image.
6, repeat steps 1-6 (I recommend using a Photoshop Action for this instead of sitting there clicking the same buttons 100 times).
The next thing that will help the effect is fixing the effect in After effects. Don't use z-space, but use scaling instead. You can parent all 100 layers to a null object and animate the scale of the null, which would apply the same speed of scale to every image. (look up "earth zoom" or "map zoom" tutorials to achieve this exact effect). With that many layers, though, you'd likely run into image buffer issues pretty quick.
Alternatively, you could apply a simple two-keyframe scale animation to one layer, select the keyframes, right click, and choose "exponential zoom." Then copy/paste the keyframes to the rest of the layers. To simplify this, stack all layers on top of one another at the beginning of the comp, paste the keyframes to all of them at the same time, then use the "sequence layers" feature to spread them out over time.
Fun video. It's nice to see a bit of how you do your work. That AI graphics stuff is pretty amazing.
To smooth the expansion, make sure that your seed images are exactly the same size and perfectly centred. Expand your first image by 200%.
Start your zoom out, with your overlay next image expanded at 200% ready to CUT (not fade) in over the old when the old hits 100%. That way you might reduce the jittering you experienced in your first attempt.
Damn it, I wish TH-cam had a backward play option, the Zoom IN would have been epic!
That's what I was hoping, that Joe's zoom out, was played in reverse, ie zoom in, across the lanscapes, ending up at the start.
@@BryTee Yeah, I think Joe needs to do a follow video which is the zoom out & then zoom in :)
Might have to make that happen.
I Zoomed IN 100x on Photoshop! (Commenter suggestion)
@@joescott Yup, that was as epic as I thought (other than the portrait orientation! ;) - Is that a limitation of a short, out of interest? (Thank you for taking the time to make that btw :D)
I was a concept artist for a small games/film company. Once Ai came in I was fired. I don't blame Ai, as such. More that my value was seen as worthless by my superiors the moment they saw the potential.
Then they immediately hired another concept artist when AI became too frustrating for them to use and couldn't give them the specifics they wanted.
An interesting alternative to what you did would have been to write a script to shrink your image by 2px*2px leaving a 1px border around the canvas, then generatively filling that, saving each image as you go. You wouldn't want to do it by hand, but if you left a PC to do this for a few days you'd end up with a crazy animation I think! Increase throughput by increasing margin size, but I think this might sacrifice some quality as it tries to fill the margin with objects from its library rather than creating a 'best fit' pixel giving what's already there.
i saw someone do this with the windows xp background, and it was.. beautiful..
6:14 "Dehance.... Dehance......Dehance...."
The stutter is absolutely fire. Great combo with the music 👌
I like the "wobble" effect very much in the video. I think I like the unintended effect better than a smooth zoom-out.
not me with a massive smile on my face as the zoom out played. Awesome video Joe! Love it.
I did an infinite zoom thing with an ai like a year back with keyboard cat, was very cool!
AI audio - did 1960s artist Tiny Tim singing "Fairies Wear Boots" by Black Sabboth with ukalalees. LAFFS BIG LAFFS!
14:40 This is what GeoGuessr pros see in their dreams
19:17 "Is it just training off of itself at this point?"
Whenever it gives you three options to choose from, I suspect the choice is fed back for further training
And so might be the usage (versus lack thereof) of the generated stock images
I respect your standpoint towards the topic 👏(from my experience, most of the guys in videos about using AI are this unempathical type of techbro who thinks everything is better if it's done by a computer... it's nice to have somebody acknoledge the harm caused by AI while still making a point about some benefits)
I like the perspective change on the little rock outcropping in the waves at 15:06 all previous images were on that little rock... imagine how small Joe is inside the cave... like a grain of sand.
You do not need a copyright holder's permission to look at a copyrighted work. You need their permission to create and distribute copies of it.
Yes you are correct, the way AI is trained is not what many people without knowledge about ML and AI think. It's more like using the data temporarily and then you won't need it again. And it's not like storing the images you train into the model. Also the way it is trained (diffusion model) makes it transformative which makes it not violate copyright law.
Stealing image in world is computer is weird. When you think about it, when you steal someone's image, you are really just copying the ones and zeros that they made.
Use this comment if you in fact, did not know.
I did because I use Photoshop and Adobe is REALLY pushing it, but it’s nice that it’s an integrated tool for when you want it
Never used photoshop, but everything seem like it has ai now
I didn't know
I didn't know.
I knew bc I pay attention to these sorts of things
As an artist who has used all traditional media, especially airbrush for over 3 decades, including Photoshop since the early 90's, I have no intention of using Ai features and have avoided even checking them out. I work on my pieces down to the pixel with brush and pen tools as though I'm using what I call analog techniques which is how I differentiate traditional methods from the digital. To me it's just a cheat, as for me "Artists that can, make art while those that cannot, use Ai".
Yep. It's wild how generative AI is readily accepted as a new art-making method. It's the easy way out.
Prepare to lose a lot of new clients since now everyone's gonna expect the results a lot faster.
This is just plain old regressive thinking. The same kind of thinking that complained about the printing press, and telephones, and TV, and computers, and internet... Try and be a bit more forward thinking.
For a long time photography itself was considered cheating when compared to painting, and then photoshop was considered cheating by photographers. In reality, they're all just tools, and so is AI.
It's probably the same with what those portraits painters thought in the early days of photography.
I got an ad for adobe, advertising its ai feature, right as he started talking about it
Next time you wanna do stuff like the edit, I recommend using a 3d program, like blender. There, you can just scale the elements based on math, move smaller images up a bit and zoom in using a linear curve.
If you truly made the images 300% bigger, you can just scale each image plane with the formula 1/3.
Importing the image planes requires an inbuilt but by default disabled addon, "images as planes"
Another easier way to do this, would be using stable diffusion and the deforum addon, which has a zoom function. it will alter the original frame, but it is basically how all those trippy zoom videos were made.
Me omw to travel from 15:28 to your house
I can’t honestly say that any of this stuff shows any signs of intelligence, artificial or otherwise. The photo edges are expanded using an algorithm that essentially copies what is already in the image. Later, larger, expansions are simply matching the image content as closely as possible to another image in a stock library of images. Given that the majority of images in photo libraries are landscapes of one sort or another, it is to be expected that all images will eventually expand into a landscape. So what we have is a complex algorithm and a huge library of images but certainly no intelligence necessary to produce these results.
In other words, exactly what I would do, given the time, finesse, patience, and catalogue of images.
Robots with AI are going to build an equivalent to the Great Wall of China at some point. Then people like yourself will argue that building structures isn't a sign of intelligence. Do note that there's also a vast catalogue of city images, and the AI did not flip over to it at any point in the zoom.
Me watching the edited sequence of photos: awesome editing, love it!
Me looking at my bed after the sequence finished: omfg eyes wait wtf you doing
Initial thoughts:
If you are getting a slight distortion on the edge of the images, make the images oversized and crop them.
You are scaling the images to 1/3rd during the generation. If you scale them back 3x in Premiere/AE and then overlay the original image in the middle, you should avoid the non-linear scaling issue.
You could nudge the generative fill a little by pushing the image a few pixels off centre and telling it to rotate up etc You could potentially get a zoom out to space if you wanted.
Edit to add... motion blur would make the zoom better imho once you fix the non-linear zoom issue.
You can expand the canvas while holding ALT + Shift, it will expand from the center, so you don't have to move the image to the center every time
Nice. I love the last few frames where it turns into a corrugated Sheetmetal.
There's a way easier method to pull this off in After Effects without getting into 3D space. Stack all your images, create a null control layer (start with the scale at something like 10,000) and parent the first image to it, scale down (on the null layer) until the first image fits within the second image (you may need to shift the position of this second image to line them up) and once they're lined up parent the second image to the null layer. Repeat on and on. And when you run out of scale margin (you will be scaling from 10,000% down to 1%) just create another null layer to control the first.
Wow, after watching the 'Results' and then looking at an object on my desk, it looked like it was expanding too! Crazy.
There is a simpler way to do the editing on After Effects. But you'd have to work from the last pictures to the first before reversing the shot.
Add your last picture to the composition, put a keyframe on scale on the first frame, then move to, let's say 2 seconds later, and scale your picture to 300% (it will add a keyframe). Then add your second picture to the composition and just parent it to the previous one. move to 2 seconds later, the scale of your first picture/layer must now be 900% (if it's a pain in the *ss to calculate each time, you can just take the scale as it is and add *3 just after, it will automatically calculate it), you add the third picture and parent it either to the second or the first (the result will be the same), then move to 6 seconds, scale your first picture to 2700%, add the fourth picture, parent it to any of the other pictures, move to 8 seconds, scale your first picture to 8100% and so on... You will then have a perfectly constant speed zoom in your creation. It might be heavy on ram since, in the end, you will have 100 layers, but it should work thought it can be optimized.
After you've done all of this, if you want a smoothest animation, you can just delete all the keyframes that are located between the first one and the last one, then select the two remaining ones and press f9 to activate smooth interpolation between keyframes, you could even just reverse the two keyframes at this point so you would zoom out instead of zooming in, just make sure to don't do that before you have added all your pictures.
Also, in premiere, while the keyframe interpolation is on smooth by default, you can have them be set to linear, but i'm not sure how it's done though, maybe by pressing ctrl while you click on the keyframe symbol... or maybe by pressing f9... i'm not as familiar with premiere as i am with after effects.
Sorry for my clunky english, i'm french. Hope it can help you with future projects.
The point about the Ai training off of itself is something my old design mentor mentioned to me recently. He had been using it as a shortcut for lower paying clients or people with lower standards, and he had realized that it had been spitting out worse and worse quality. He also that more and more prompts were being censored, which would put restrictions on what it could train off of or generate affecting the quality. However it shouldn't affect it as much as the quality had been showing, meaning it probably has been training off of itself, creating a negative feedback loop.
I am a huge Adobe nerd, after effects being my favourite! I love your channel, such a treat to have an episode about Adobe hell yeh
Everytime I've seen this expanding image effect, it had this wobble between transitions. In any case, your friend did a great job with this weird AI journey.
with the editing issue; you could've also just placed all the images in their right places, and anchor them to each other. scaling the bigger one would've made the smaller one follow right along.
JEZUZ CRUST!!! After watching the edited version my room was literally distorting before my very eyes when I took my eyes off the screen!! Talk about trippy!
Good job matching the stutter with the beat of the music.