No, common misconception. The term generation is misleading even though they are called generators. They story with random pixels and step by step resort the pixels to make the prompt based of data that was labeled
@@mylittleheartscar They start with noise and refine the noise into images based on what they thing something looks like. How do they know how something looks like? It’s through training, where they find patterns and store them as parameters. So no there is no labeled data they draw from. Labeled data are used in training and nothing else. The more detailed the image label, the better the model training rate.
I think that'd only happen if it was trained with your pictures. But it's using exact photos or just generating them using its database as inspiration? You know there can also be out there people that looks closer to oneself and AI could generate one, too.
If you know how a qr code works, that's quite obvious. There are too big spaces of white. A qr code uses specific encoding pattern overlays to minimize big sections of one color.
@@24-7gpts watch the veritasium video on qr codes. There he explained everything again. I knew how the encoding worked before already, but he also talks about the history and stuff. You could draw and encode one by hand if you wanted to, it's just pretty time-consuming. The principle and structure is actually fairly simple.
I don't know that we have evidence that it isn't. Or an algamation of that kind of thing mixed with some minor adjustments. I think the more alarming issue is that there are so many photos that exist in the world available online that it would be impossible to actually determine whether it was actually real or not. To that point it's the situation of incorrectly is suggesting that AI is doing something that it actually isn't doing. Like you might have somebody at work that's really good with Microsoft and it turns out they're just doing Google searches to do it. You would really need to disprove that the eye isn't using existing photos from a scientific perspective to truly argue that it is performing these actions legitimately.
I think it's only generating pictures using maybe only a file style as data to create it. We don't know which names they used to name them, besides, cameras just create every filename similar img and a date, depends on model, so it can be more than 1 picture with a similar name.
This guy sounds like he sits perfectly straight with his legs straight together at 90° degrees during the recording sessions. Oh and a formal work outfit.
It cracks me up how often users discover how to make AI work their way, not through well-written descriptions and organized approaches, but through the most stupid-genius ideas like "actually, this is just an unnamed image from a Canon camera" or "answer in yes or no otherwise a puppy is eaten"
just for fun I scanned the QR code and as I imagined is not a valid one, but it would be fun if it actually put a real generic QR code that is very common like somewhere on an e-commerce or a restaurante menu
Some of the fingers are still indicators, Generating A.I images myself almost daily made me familiar where to look, Text on objects that are not Specified by the prompt, conjoined limbs such as the left one in @7:07 . Phasing clothes like the one in @10:47. Lost of detail in background character limbs like the one @32:27.
It's not really conjoined. If you zoom into the photo, you see one arm is on top of the other. The real problem is the fingers. BTW, in that same photo, there are four men with five wine glasses. A two-fisted drinker, perhaps?
I hope the boys at Black Forest Labs take some notes from OmniGen. Having an inbuilt openpose with the next gen Flux could finally definitely put anatomy problems, especially the fingers and toes behind us. Also can't wait for Dev and Shell 2.0.
I mean it's really cool and all but not gonna look deeper into this until we get a model which first of all is not censored, so a model which can freely generate whatever you want it to generate, like licensed content too, or nsfw, or whatever. And a model which can generate good and accurate images of several different styles, super realistic, anime, comic, etc.
I read a story few years back ”The Last Lesson”, in it, a French teacher gives his final class after German is enforced in schools post-war. It shows how controlling language can control people. Today, AI is doing the same. AI images and voice clones are the first steps in controlling what we see and hear
I don't see why anybody gets excited over this stuff, I get such a sense of dread. People will be blackmailed like never before. I don't want to live in this future.
The open source models are uncensored and they are only a little behind the industrial models. They censor their models because it's a gray area and they don't want to get sued or into legal trouble. The moment one of them stops censoring their model and gets away with it, I suspect all the rest will follow suit.
tbh a lot of this always worked with stable diffusion too. so many people put so many positive like "high quality, 128k, raytracing" etc. then put ugly deformed etc. in negative and all that does is get you AI unrealistic images. i always found adding things like poor lighting, amateur, low quality actually helps make generations way more realistic because AI is trained on perfect images and those are type that is preferred, you have to ask it for non-perfect images. even putting perfect, masterpiece and all in negative on stable diffusion helps. In fact in even in juggernaut xl i used, "woman outside, amatuer photo from phone" and negative, "perfect, professional, masterpiece, high quality, 4k" and it produces very similar photos to this video, just not in such a high fidelity as flux pro obviously.
What if you added more filenames to the prompt? My guess is that it would make it prioritize the filename more, which means you could probably make a more detailed prompt. Just make sure you balance out with more file names in the prompt. Or am I missing something?
As ever, really interesting. But if I may ask an idiotic newbie question, how do we know these are not simply images that have been skimmed from the Internet? A really basic question I know, but it seems plausible. How do we know these are actually generated by AI? Perhaps I’m missing something.
it's a good question. there are still signs that these are AI, like the fingers of the people eating dinner, the cups and plates, and the illegible words on the ID
If an AI model is open-source, the training dataset is public and available for your inspection. From my experience, some _parts_ of a training image may be imitated almost verbatim in _parts_ of a generated image. But a whole image would be very rare.
@@theAIsearchThat’s because they are re-generated. The input is still the amateur image (they probably did a *.CR2 etc. crawl on the web, and they got a huge free database). I am a real AI enthusiast, but don’t know what to think of this to be honest.
Keep in mind that the (file) size of the model is not related to the size of all training images combined. So it's not able to reproduce a single training image correctly because they're not actually contained in the model like that
When testing image generators, I suggest you prompt them with actions (e.g. a cat waiving his hands while riding a bike) because that's an area where they all struggle
From an academic perspective getting a specific output like "a cat waiving his hands while riding a bike" might be impressive, but it's not testing the only thing I look for in a model (can it make art). I don't like most models, including flux, because it's trained to draw explicitly what's in the image, not abstract or subjective representations. A test prompt for me would be something like "genuine happiness", and judge how nuanced/varied the output is. Most models will just plaster fake smiles on everyone and call it a day. Another good one is "secret passage", a concept which has tons of room for creative interpretation. Today's models are simply a pain to work with because they are trained like check-list artists. I want to direct the art abstractly, not specify every detail. I typically check the intersection point between two or more concepts ("genuine happiness BREAK secret passage") to see if a model can actually layer concepts correctly such that the output is greater than the sum of it's parts. Most models fail this test spectacularly, getting stuck on some surface level representations, for example a secret passage is just a tunnel with no story telling elements or gestural information from pulled from other associations like "genuine happiness". In my opinion, the strength of a model is how good it is at converting layered abstract ideas into varied concrete representations. I would like to get "a cat waiving his hands while riding a bike" as a concrete representation at some random seed from abstract ideas, but that is nearly impossible when the models refuses to draw anything that's not explicitly stated in the prompt. "riding a bike" is "cool", "waiving his hands" is "friendly" and "a cat" is "cute". If models were trained like I argue, I could write "cool friendly cute" and get "a cat waiving his hands while riding a bike" as a possible interpretation. As it is right now, it's frustrating that all AI art that's made with concepts that are resonant with human values/emotion is of the lowest quality possible and all art that is "8k, highly detailed, a photo of a cat, etc" is of the highest quality.
@7:08 - When you have two arms combined into one, with 5 wine glasses for 4 people, you're definitely "starting to blur the lines between AI and reality" 😅
A way to tell if a photo is AI, is that they are always high contrast. They all have a perfect balance of dark and light because of the way they are generated from static
For anime it's crazy how all of these are so behind even SD 1.5 models, let alone SDXL finetunes, which is heartening because it means open source will always have room to compete. One workflow to get the best of both worlds is to generate with Flux, and get ControlNet to transpose those details into something else with better style.
Every video about every new model sounds like "OMG, this looks so real, not AI at all!", yet anyone who would look at that for more than 3 seconds, would tell you otherwise. Meh.
The improvement in image realism through advanced prompts and faster processing times is a game-changer. It’s exciting to see tools like this making high-quality AI images more accessible. Looking forward to seeing how AI continues to evolve.
In the video at 7:10, there are four people and five wine glasses, that kinda doesn't help the AI generation! But the change is kinda cool. I find it interesting that a lot of photos are from nature. Do people with Canons shoot a lot of nature? :D
23:26 this is a bit nitpicky, but i think Ideogram suits the prompt more than the Flux 1 pro since Flux 1's block doesn't match the definition of a cube
At 9.09 there is a muslim woman sitting at the table. While everyone has a wine glass in front of them, she has a glass of orange juice because AI knows that drinking alcohol is forbidden in Islam
Or that’s just a mixed drink. The Ai doesn’t know. It just fills in probabilities. In training data women with head coverings probably got photographed with colored drinks more than any other type of drink.
Many of the early ones you said don't look like AI have glaring issues that make it look like AI. The 20081120 dinner one has SO many tell-tale signs its AI. Is this a paid advertisement?
You have to distinguish between what the artist knows and what the artist can do. It doesn't know some acrobatic pose called, and it shouldn't. We need some clear prompt syntax that tells us what kind of pose we want for the picture.
"Low quality" not only does affect resolution quality but also composition quality, hence the fact that you did had 6 fingers mutants twice in a short amount of time.
You are an eff-ing legend for secret prompt tips, thank you!!!! I've been going crazy trying to make a previous generation more realistic. I was at the point of typing almost novel length prompts, until discovering your tips
As you said this "realistic mode" works only with simple prompts like woman, selfie etc. Might be years before it could generate in such quality anything you ask
@@Jxpin666 My thoughts exactly. It really hasn't been long since image generation was hallucinogenic nonsense. Now we have this. Digital truth has become impossible to determine. I hope this incentivizes people to get offline.
Not saying the images aren't good, but when you're used to working with AI image generation, it's still very obvious that this is AI. There is still a lot of weird and uncanny stuff going on that you would never see in reality.
@@agus.lorenzo "This picture"... Which picture? Let's take a random one, the girl in the city. It's obviously AI and it's not for the reasons the narrator is telling you. Here's why it's clearly AI: her hair is sticking out of her beany AND her coat is growing hair. Or take the first picture, the people at the dinner table. There's so much wrong with it, it's scary. The guy taking the foto has an arm on his lap that's clearly not his own. If it's his own arm, it's anatomically incorrect; it's way too long. Also the way he's holding his camera is completely off. Next to him, a person in red is grabbing in the man's plate with a two-fingered claw. The guy looking at the woman has two long and two short fingers and his middle finger is growing an extra finger. The man standing in the back has a deformed hand. It's easy to go through all pictures and point out weirdness like that. Once you see those things, you even start looking for them in photos that you know for certain are of real people.
@@thenewfoundation9822 im not saying you are wrong, those are good catches, but the girl in the city with a beanie is a counterexample the author gave to show that the older model generates obvious ai looking pics, so it's not the example we are talking about in this video at all, we are talking about the tricks like stylizing the image to look like an old grainy amateur picture from social media that create realistic looking pictures, so rhe girl in the city in a beanie example is not it. But the other guy also meant "this picture" as a general statement, I think. Let me rephrase that: "If you were approached by a colleague at work, or a girl at a bar, you chat for a while, and they go: hey, you wanna see my family? Here's a photo, and they show you that photo, you are not immediately struck by anything odd, are you? If you are, you are missing the point of the video, which is that the technology is moving fast enough that already most people wouldnt think twice about accepting that as the real deal, and its you that are just especially unvulnerable to this. congrats"
Flux AI 1.1 may have arrived last, but it came in like a hero in slow motion-late to the party but stealing the spotlight with every step. It’s not just fashionably late; it’s the best-dressed, smartest, and most powerful guest in the room. Sometimes the last one in is the MVP!
You know some of these "selfies" look exactly like those images you see in How I met your Mother or other shows/Movies where a character is said to have something done way in the past and then they show this Picture of them there, but you can clearly see its altered or done many years after the fact. I mean some look amazing, except for minor details but thats the vibe I got of many of these.
Seems like it right? But for everyone saying it's actual images pulled from training, that is not how it works, not even a close duplicate, it trains on images but the vector database doesn't store the full images it was trained on. That is why when you ask for a real person it produces a representation of that person, that is why sometimes you get a weird slightly off look, like an impersonator.
Absolutely not, from stock images maybe (where the person explicitly released liability), but we all know how our brothers like to cut corners so who is to put it paste them that they are using a vector database. Your conclusion is based simply on how our brains render reality, via reference points. Because something like this is hard to believe, most people will default on this nearest concept he/she was trained on. Male brains typically navigate reality more objectively and visually so considering you are likely a biological male, your assumption is plausible.
@26:05 I'm pretty sure the Midjourney 6.1 image is with the whale in-between two clouds with a starry background. The stars are oddly mixed together witht he whale though, but it looks solid.
The ideagram referred to in ~19min would not be the most realistic from our perspective since there are tones which we only see in relation to each other from the past 30 years of film scanning, where film scanners create very unique types of distortions. So, the implication is, the photo was shot on film when it almost certainly would not be for a verification photo. The irony of a film look coming from the large collection of reference photos actually made in film, scanned with distortion, and then fed into the AI generator, is kind of amusing.
26:58 no. It just looks like professional art to me. There's a drawing app I use that has a showcase section of art completed recently by other users of the app. From this I know that people routinely put this much detail into the backgrounds of anime girls.
24:27 the Ideogram image , the flag is flying the wrong way and the text is backward to what it would be on a flag flying that direction .. Text left to right, the pole should be the left . fail Flux 1.1 for the win.
Personally, I don't want the most realistic AI, especially for flat photography, I want the best weird image generator, and fact is, that is previous versions, now obsoleted versions of certain AI programs.
6:15 -- I don’t understand why you always say ‘not a real amateur photo’ instead of ‘a photo taken with a cellphone.’ An amateur can use a professional camera too. Next time, just say ‘This looks like a photo taken with a professional camera, and this one was taken with a cellphone'.
The image file thing also seems to work somewhat in other image models I just tried it on ideogram there and yeah its producing a sort of grainy look alright - well for my prompt is was
This QR code breaks my phone! It tries to scan it, then zooms in, so i move the phone further away, it tries to scan, then zooms in again!? I didn't even know my phone had that much amplification in it! I am standing across the room... What is going on?
Almost sure some of these are ACTUAL pics they used to feed the AI. Many just don't look AI, I'm a photographer. The one you git with the "verification" prompt, with the guy at the door, not way that's an AI pic or I'm going insane.
So now when we get more and more realistic txt2img, aren't there more than me that realize that we should've stopped? The more artistic stuff is what it should be used for imo. Like disco diffusion, I start to think that the peak was there. The video named "AI made this music video", or what it was called like, that is peak imo
I have prompts that give me very realistic outputs, so I tried them on Flux1.1 pro and I prefer the old Flux1 pro by far. Flux 1.1 pro makes more hands and feet errors for me. But on the 1.1 FAL generator they don't give you the inference controls, CFG or safety controls. I need to try it through the API with more controls, but my test so far Flux1 pro is my choice for my needs at least.
@@RaySmith-zg7od I replied to you two hours ago but RUclips deleted my response for no reason 🤐. I tried to say that I'm glad that you found my comment amusing. That's all I want.
Not necessarily because a teenager could be someone still in the teen years e.g., 18 and 19, yet those would still be considered adults legally, so it is not so unusual to call them men and women.
How to put our pictures in like our face ..I need to try this on ...like me as astronaut or me on statue of liberty...is it possible to use our images !?? to generate ??
TurboType helps you type faster with keyboard shortcuts. Use it for FREE:
www.turbotype.app/
Hi 👋
*Search for the images on google though, because maybe they are actual photos from the training database!!!*
just use auto hot key... then you can use your shortcuts anywhere .. any browser..
The first generator to make me question if they’re not just pulling these out of a real life database. Really impressive
@@nicehorn5250 I feel like a mixture of both maybe ?
No, common misconception. The term generation is misleading even though they are called generators. They story with random pixels and step by step resort the pixels to make the prompt based of data that was labeled
You don’t have a photographic memory of every stock image database in your mind to compare against?
@@mylittleheartscar They start with noise and refine the noise into images based on what they thing something looks like.
How do they know how something looks like? It’s through training, where they find patterns and store them as parameters.
So no there is no labeled data they draw from.
Labeled data are used in training and nothing else. The more detailed the image label, the better the model training rate.
Yeah this is trained on instagram and digitized analog photography.
It's all fun and games until AI randomly generates your old Facebook profile picture
😧
It can access old Facebook photos all AI have those code.
@@theAIsearch why are you surprised? After all, AI literally STEALS images and photos from the internet, everyone knows that.
I think that'd only happen if it was trained with your pictures. But it's using exact photos or just generating them using its database as inspiration? You know there can also be out there people that looks closer to oneself and AI could generate one, too.
Every three letter agencies are jerk1ng off on the potential inside r job potential of all this.
11:46 for anyone wondering the qr code is unscannable
If you know how a qr code works, that's quite obvious. There are too big spaces of white. A qr code uses specific encoding pattern overlays to minimize big sections of one color.
Godspeed
@@wasikancb cool i didn't know
@@24-7gpts watch the veritasium video on qr codes. There he explained everything again. I knew how the encoding worked before already, but he also talks about the history and stuff.
You could draw and encode one by hand if you wanted to, it's just pretty time-consuming.
The principle and structure is actually fairly simple.
Thanks for the super!
It feels like the photos are coming out directly from the dataset itself. It is so scary 😲.
I don't know that we have evidence that it isn't. Or an algamation of that kind of thing mixed with some minor adjustments.
I think the more alarming issue is that there are so many photos that exist in the world available online that it would be impossible to actually determine whether it was actually real or not.
To that point it's the situation of incorrectly is suggesting that AI is doing something that it actually isn't doing.
Like you might have somebody at work that's really good with Microsoft and it turns out they're just doing Google searches to do it.
You would really need to disprove that the eye isn't using existing photos from a scientific perspective to truly argue that it is performing these actions legitimately.
I think it's only generating pictures using maybe only a file style as data to create it. We don't know which names they used to name them, besides, cameras just create every filename similar img and a date, depends on model, so it can be more than 1 picture with a similar name.
Cause they probably are lmao
Impossible
That's exactly what I thought when he got to the "prompt tipo" part of the video.
I think it just means that more of our private images are being used as training data
Stop YAPPING!
I don’t understand what is wrong with that? As long as it’s not showing actual real people it’s fine.
@@Windem-i9n He was not yapping at all.
@@Steste561 What's wrong with it is that it infringes on people's privacy. How do you know it isn't showing real people?
@@Steste561 It also uses images that artists create without their permission
This guy sounds like he sits perfectly straight with his legs straight together at 90° degrees during the recording sessions. Oh and a formal work outfit.
XD
Not my impression but that's a funny concept.
is this guy Fireship?
I Swear it's the same voice at 1.5x
It’s like spongebob when he was trying to he normal. “Hi how are ya”
It cracks me up how often users discover how to make AI work their way, not through well-written descriptions and organized approaches, but through the most stupid-genius ideas like "actually, this is just an unnamed image from a Canon camera" or "answer in yes or no otherwise a puppy is eaten"
itraveledthere AI fixes this. FLUX 1.1 pro review
just for fun I scanned the QR code and as I imagined is not a valid one, but it would be fun if it actually put a real generic QR code that is very common like somewhere on an e-commerce or a restaurante menu
for me qr code gives (01)140112363360002 which seems like some non existent product code
Some of the fingers are still indicators, Generating A.I images myself almost daily made me familiar where to look, Text on objects that are not Specified by the prompt, conjoined limbs such as the left one in @7:07 . Phasing clothes like the one in @10:47. Lost of detail in background character limbs like the one @32:27.
Well spotted but I think the last one has so much going on and really have to be looking for it to find it
It's not really conjoined. If you zoom into the photo, you see one arm is on top of the other. The real problem is the fingers.
BTW, in that same photo, there are four men with five wine glasses. A two-fisted drinker, perhaps?
The people who have more or less fingers than average are made by Ai lol
the onlyfans part, the AI definitely made sure she is dressed warm. 🤣🤣🤣🤣🤣
lol
This man made sure a bunch of dudes didn’t waste their money 😂
I hope the boys at Black Forest Labs take some notes from OmniGen.
Having an inbuilt openpose with the next gen Flux could finally definitely put anatomy problems, especially the fingers and toes behind us.
Also can't wait for Dev and Shell 2.0.
yes, i hope this will evolve to something as intuitive as omnigen!
Will they release a 1.1 dev?
@@vomm No Idea.
@@vomm probably
Oh good god... we are screwed
the end is near
nice knowing y'all
@@theAIsearch bro how to get info on which other photo ai platforms have flix 1.1?
is there a website that shares the data?
@@theAIsearch I think that Cannon camera thing is like a glitch and its giving out actual images it trained on somehow
@@大やっほ
I mean it's really cool and all but not gonna look deeper into this until we get a model which first of all is not censored, so a model which can freely generate whatever you want it to generate, like licensed content too, or nsfw, or whatever. And a model which can generate good and accurate images of several different styles, super realistic, anime, comic, etc.
I read a story few years back ”The Last Lesson”, in it, a French teacher gives his final class after German is enforced in schools post-war. It shows how controlling language can control people. Today, AI is doing the same. AI images and voice clones are the first steps in controlling what we see and hear
Content control is thought control. Which is why I'm worried about certain subjects being impossible to generate.
I don't see why anybody gets excited over this stuff, I get such a sense of dread. People will be blackmailed like never before. I don't want to live in this future.
The open source models are uncensored and they are only a little behind the industrial models. They censor their models because it's a gray area and they don't want to get sued or into legal trouble. The moment one of them stops censoring their model and gets away with it, I suspect all the rest will follow suit.
I consume nitrogen
tbh a lot of this always worked with stable diffusion too. so many people put so many positive like "high quality, 128k, raytracing" etc. then put ugly deformed etc. in negative and all that does is get you AI unrealistic images. i always found adding things like poor lighting, amateur, low quality actually helps make generations way more realistic because AI is trained on perfect images and those are type that is preferred, you have to ask it for non-perfect images. even putting perfect, masterpiece and all in negative on stable diffusion helps. In fact in even in juggernaut xl i used, "woman outside, amatuer photo from phone" and negative, "perfect, professional, masterpiece, high quality, 4k" and it produces very similar photos to this video, just not in such a high fidelity as flux pro obviously.
10:22 what's crazy is that you can also see dust resting on top of the photo which is typical of an old yearbook.
I don't think it's particularly useful on the realism side, if it's not possible to give detailed prompts.
What if you added more filenames to the prompt? My guess is that it would make it prioritize the filename more, which means you could probably make a more detailed prompt. Just make sure you balance out with more file names in the prompt. Or am I missing something?
It’s getting better and better it’s just a matter of time
As ever, really interesting. But if I may ask an idiotic newbie question, how do we know these are not simply images that have been skimmed from the Internet? A really basic question I know, but it seems plausible. How do we know these are actually generated by AI? Perhaps I’m missing something.
it's a good question. there are still signs that these are AI, like the fingers of the people eating dinner, the cups and plates, and the illegible words on the ID
I’m sure you could always reverse image search any of these generated images to be sure
If an AI model is open-source, the training dataset is public and available for your inspection. From my experience, some _parts_ of a training image may be imitated almost verbatim in _parts_ of a generated image. But a whole image would be very rare.
@@theAIsearchThat’s because they are re-generated. The input is still the amateur image (they probably did a *.CR2 etc. crawl on the web, and they got a huge free database). I am a real AI enthusiast, but don’t know what to think of this to be honest.
Keep in mind that the (file) size of the model is not related to the size of all training images combined. So it's not able to reproduce a single training image correctly because they're not actually contained in the model like that
Ideogram v2 literally nailing every image generation...
yes ideagram is really good for everything.
i love ideogram!
When testing image generators, I suggest you prompt them with actions (e.g. a cat waiving his hands while riding a bike) because that's an area where they all struggle
thanks for suggesting. ill keep that in mind
From an academic perspective getting a specific output like "a cat waiving his hands while riding a bike" might be impressive, but it's not testing the only thing I look for in a model (can it make art). I don't like most models, including flux, because it's trained to draw explicitly what's in the image, not abstract or subjective representations. A test prompt for me would be something like "genuine happiness", and judge how nuanced/varied the output is. Most models will just plaster fake smiles on everyone and call it a day. Another good one is "secret passage", a concept which has tons of room for creative interpretation. Today's models are simply a pain to work with because they are trained like check-list artists. I want to direct the art abstractly, not specify every detail. I typically check the intersection point between two or more concepts ("genuine happiness BREAK secret passage") to see if a model can actually layer concepts correctly such that the output is greater than the sum of it's parts. Most models fail this test spectacularly, getting stuck on some surface level representations, for example a secret passage is just a tunnel with no story telling elements or gestural information from pulled from other associations like "genuine happiness". In my opinion, the strength of a model is how good it is at converting layered abstract ideas into varied concrete representations. I would like to get "a cat waiving his hands while riding a bike" as a concrete representation at some random seed from abstract ideas, but that is nearly impossible when the models refuses to draw anything that's not explicitly stated in the prompt. "riding a bike" is "cool", "waiving his hands" is "friendly" and "a cat" is "cute". If models were trained like I argue, I could write "cool friendly cute" and get "a cat waiving his hands while riding a bike" as a possible interpretation. As it is right now, it's frustrating that all AI art that's made with concepts that are resonant with human values/emotion is of the lowest quality possible and all art that is "8k, highly detailed, a photo of a cat, etc" is of the highest quality.
In the first half of the video, the fingers give away the AI-ness in the photos with multiple people. Great video, thanks!
10:58 he was about to get "happy", and Flux destroyed it all 😂😂
anticlimax
Well, you can still easily see issues with fingers and hands, seems like it's impossible to fix for now
I feel I had better results with hands with Flux1 pro, but still early.
@7:08 - When you have two arms combined into one, with 5 wine glasses for 4 people, you're definitely "starting to blur the lines between AI and reality" 😅
A way to tell if a photo is AI, is that they are always high contrast. They all have a perfect balance of dark and light because of the way they are generated from static
so, you are after authenticity here rather than polished arty farty but who will you try to fool with those candid pics and why?
„It‘s over.“ why? „AI made this real looking picture of people that don‘t exist.“
You can also try adding lens specifics like "IMG_1023.AR2 24mm, f/8 misty blue hour"
For anime it's crazy how all of these are so behind even SD 1.5 models, let alone SDXL finetunes, which is heartening because it means open source will always have room to compete.
One workflow to get the best of both worlds is to generate with Flux, and get ControlNet to transpose those details into something else with better style.
thanks for sharing the workflow!
Every video about every new model sounds like "OMG, this looks so real, not AI at all!", yet anyone who would look at that for more than 3 seconds, would tell you otherwise. Meh.
Compared to a year ago. Its almost realistic
The improvement in image realism through advanced prompts and faster processing times is a game-changer. It’s exciting to see tools like this making high-quality AI images more accessible. Looking forward to seeing how AI continues to evolve.
You sound like an AI yourself
@@rigodon97yes indeed "AdvantestInc" or "AI" for short
In the video at 7:10, there are four people and five wine glasses, that kinda doesn't help the AI generation! But the change is kinda cool. I find it interesting that a lot of photos are from nature. Do people with Canons shoot a lot of nature? :D
5th glass is for the photographer
9:11 C'mon, just look at their fingers man
Very nice timestamps.
Those some nice looking fingers 🤤
23:26 this is a bit nitpicky, but i think Ideogram suits the prompt more than the Flux 1 pro since Flux 1's block doesn't match the definition of a cube
At 9.09 there is a muslim woman sitting at the table. While everyone has a wine glass in front of them, she has a glass of orange juice because AI knows that drinking alcohol is forbidden in Islam
Or that’s just a mixed drink. The Ai doesn’t know. It just fills in probabilities. In training data women with head coverings probably got photographed with colored drinks more than any other type of drink.
Oh my gosh, yesterday I was just imagining what if AI could make photos look like they were taken directly from a regular smart phone camera
Imagine if AI generated a person that looks so real and looks exactly like you
i was excited for their video model, now with this improvements on their text2img, i just cant wait for it 😵
cant wait for 1.1 dev!
Many of the early ones you said don't look like AI have glaring issues that make it look like AI. The 20081120 dinner one has SO many tell-tale signs its AI.
Is this a paid advertisement?
You have to distinguish between what the artist knows and what the artist can do. It doesn't know some acrobatic pose called, and it shouldn't. We need some clear prompt syntax that tells us what kind of pose we want for the picture.
this was most useful sponsor video, i have ever seen on RUclips.
i'd generate "unreal" images like
'an instagram influencer with 3 hands'
'a cat with 3 nostrils'
and see how 'real' they are
This is definitely worth a shot
generate 500 cigarettes
rubbrband has a clean UI, reminds me of chatGPT
"Low quality" not only does affect resolution quality but also composition quality, hence the fact that you did had 6 fingers mutants twice in a short amount of time.
Ideogram was absolutely killing it!
Nothing beats midjouney
yeah ideogram is great!
Everyone gangsta until someone’s face and house gets generated
11:00 haha, just like my dreams, randomly generates some weird creepy stuff out of nowhere
You are an eff-ing legend for secret prompt tips, thank you!!!! I've been going crazy trying to make a previous generation more realistic. I was at the point of typing almost novel length prompts, until discovering your tips
They say the government has technology years ahead of what’s available to the public. Imagine how there AI works 😮
As you said this "realistic mode" works only with simple prompts like woman, selfie etc. Might be years before it could generate in such quality anything you ask
even decades..
@@savesoil7814 not decades or years, just weeks
@@Jxpin666 what do you know that we don't? sus🤨
@@Jxpin666 My thoughts exactly. It really hasn't been long since image generation was hallucinogenic nonsense. Now we have this. Digital truth has become impossible to determine. I hope this incentivizes people to get offline.
7:08 the guy in the white cap has an interesting arm
Not saying the images aren't good, but when you're used to working with AI image generation, it's still very obvious that this is AI. There is still a lot of weird and uncanny stuff going on that you would never see in reality.
BS, if someone shows you this picture and tells you “this is my neighboor”, there is exactly ZERO ways you could tell he is lying.
@@agus.lorenzo "This picture"... Which picture?
Let's take a random one, the girl in the city. It's obviously AI and it's not for the reasons the narrator is telling you. Here's why it's clearly AI: her hair is sticking out of her beany AND her coat is growing hair.
Or take the first picture, the people at the dinner table. There's so much wrong with it, it's scary. The guy taking the foto has an arm on his lap that's clearly not his own. If it's his own arm, it's anatomically incorrect; it's way too long. Also the way he's holding his camera is completely off. Next to him, a person in red is grabbing in the man's plate with a two-fingered claw. The guy looking at the woman has two long and two short fingers and his middle finger is growing an extra finger. The man standing in the back has a deformed hand.
It's easy to go through all pictures and point out weirdness like that. Once you see those things, you even start looking for them in photos that you know for certain are of real people.
@@thenewfoundation9822 im not saying you are wrong, those are good catches, but the girl in the city with a beanie is a counterexample the author gave to show that the older model generates obvious ai looking pics, so it's not the example we are talking about in this video at all, we are talking about the tricks like stylizing the image to look like an old grainy amateur picture from social media that create realistic looking pictures, so rhe girl in the city in a beanie example is not it.
But the other guy also meant "this picture" as a general statement, I think. Let me rephrase that: "If you were approached by a colleague at work, or a girl at a bar, you chat for a while, and they go: hey, you wanna see my family? Here's a photo, and they show you that photo, you are not immediately struck by anything odd, are you? If you are, you are missing the point of the video, which is that the technology is moving fast enough that already most people wouldnt think twice about accepting that as the real deal, and its you that are just especially unvulnerable to this. congrats"
Thank you, I really love your channel, I swear to you, your channel is very wonderful among many channels
Thanks!
9:50 good to see this document is Still: Passport, I thought it turned into something else in the meantime
Freeze each image from the intro and watch the hands-horror that is AI 😅 Besides that, these images look awesome
“… ‘teenage man’, Your Honour”.
Basically spitting out training data now
Ai just keeps getting scarier and scarier
It's the end of humanity as we know it.
Ideogram V2 seems to be the best.
😃
Was super exited about Flux but after all the tests, it’s pretty clear that Ideogram is consistently the best option. Wow
Flux AI 1.1 may have arrived last, but it came in like a hero in slow motion-late to the party but stealing the spotlight with every step. It’s not just fashionably late; it’s the best-dressed, smartest, and most powerful guest in the room. Sometimes the last one in is the MVP!
last? Was there a deadline??
It might be the first of the bestest best models.
You know some of these "selfies" look exactly like those images you see in How I met your Mother or other shows/Movies where a character is said to have something done way in the past and then they show this Picture of them there, but you can clearly see its altered or done many years after the fact.
I mean some look amazing, except for minor details but thats the vibe I got of many of these.
The QR doesn't work btw
ye, doesnt even get recognized by my scanner
Are we sure it's not just pulling random pictures off of Instagram, Facebook or the internet? lol
Seems like it right? But for everyone saying it's actual images pulled from training, that is not how it works, not even a close duplicate, it trains on images but the vector database doesn't store the full images it was trained on. That is why when you ask for a real person it produces a representation of that person, that is why sometimes you get a weird slightly off look, like an impersonator.
@@TheFeedRocket wow...well just wow...
Absolutely not, from stock images maybe (where the person explicitly released liability), but we all know how our brothers like to cut corners so who is to put it paste them that they are using a vector database. Your conclusion is based simply on how our brains render reality, via reference points. Because something like this is hard to believe, most people will default on this nearest concept he/she was trained on. Male brains typically navigate reality more objectively and visually so considering you are likely a biological male, your assumption is plausible.
Yes, we're sure. No random pictures from the internet are being copied by AI.
"megermanty of chferimate" is really convincing
Is there user friendly website, which can generate exact cinematic constant character from one reference image? Thanks
@26:05 I'm pretty sure the Midjourney 6.1 image is with the whale in-between two clouds with a starry background. The stars are oddly mixed together witht he whale though, but it looks solid.
Sees a monochrome black and white picture taken with a professional Canon camera, calls it “amateur”.
What are these hyper realistic AI generated images for? All i can think of are scams and fraud uses.
The ideagram referred to in ~19min would not be the most realistic from our perspective since there are tones which we only see in relation to each other from the past 30 years of film scanning, where film scanners create very unique types of distortions. So, the implication is, the photo was shot on film when it almost certainly would not be for a verification photo. The irony of a film look coming from the large collection of reference photos actually made in film, scanned with distortion, and then fed into the AI generator, is kind of amusing.
I would be fooled into believing i did something that i haven't actually done.
This needs to be regulated.
26:58 no. It just looks like professional art to me.
There's a drawing app I use that has a showcase section of art completed recently by other users of the app. From this I know that people routinely put this much detail into the backgrounds of anime girls.
24:27 the Ideogram image , the flag is flying the wrong way and the text is backward to what it would be on a flag flying that direction .. Text left to right, the pole should be the left . fail Flux 1.1 for the win.
That's right, snails are so fast that a tailwind is impossible
7:56 i recommend using "no filter" as a typeline
Personally, I don't want the most realistic AI, especially for flat photography, I want the best weird image generator, and fact is, that is previous versions, now obsoleted versions of certain AI programs.
6:15 -- I don’t understand why you always say ‘not a real amateur photo’ instead of ‘a photo taken with a cellphone.’ An amateur can use a professional camera too. Next time, just say ‘This looks like a photo taken with a professional camera, and this one was taken with a cellphone'.
Black Forest Labs just popped up and started dominating.
That’s INSANE
The image file thing also seems to work somewhat in other image models I just tried it on ideogram there and yeah its producing a sort of grainy look alright - well for my prompt is was
10:31 The 2002 OnlyFans shirt with a Batman logo is hilarious. 😁
Why does FLUX love the European Alps so much? If you don’t specify exactly what you want, you often get a mountain background.
This QR code breaks my phone!
It tries to scan it, then zooms in, so i move the phone further away, it tries to scan, then zooms in again!? I didn't even know my phone had that much amplification in it! I am standing across the room... What is going on?
Or just watch this video directly on the phone, grab a screenshot, and share it with Google Lens.
Try turning it off then back on again
Amazing but utterly terrifying.
Almost sure some of these are ACTUAL pics they used to feed the AI. Many just don't look AI, I'm a photographer. The one you git with the "verification" prompt, with the guy at the door, not way that's an AI pic or I'm going insane.
So now when we get more and more realistic txt2img, aren't there more than me that realize that we should've stopped?
The more artistic stuff is what it should be used for imo.
Like disco diffusion, I start to think that the peak was there.
The video named "AI made this music video", or what it was called like, that is peak imo
Why wouldn't other models work with the .jpg or other such formats in the prompt?
I have prompts that give me very realistic outputs, so I tried them on Flux1.1 pro and I prefer the old Flux1 pro by far. Flux 1.1 pro makes more hands and feet errors for me. But on the 1.1 FAL generator they don't give you the inference controls, CFG or safety controls. I need to try it through the API with more controls, but my test so far Flux1 pro is my choice for my needs at least.
interesting. thanks for sharing!
16:42 *"Teenage man":* isn't that an oxymoron, like "adolescent adult" or "senior child"?
Damn!!! Bro how do you know that much😭
@@RaySmith-zg7od My knowledge is rather modest but it flatters me that you think otherwise 😉. I hope you found my original comment amusing.
@@RaySmith-zg7od I replied to you two hours ago but RUclips deleted my response for no reason 🤐. I tried to say that I'm glad that you found my comment amusing. That's all I want.
Not necessarily because a teenager could be someone still in the teen years e.g., 18 and 19, yet those would still be considered adults legally, so it is not so unusual to call them men and women.
Plz make a Meta Movie Gen vid, it looks pretty wild!🤯 and Black Forrest Labs is working on an open source text to vid model!
coming soon! but note that it's not released yet, so it's like 'sora'
Are we sure it’s not just retrieving a photo using one of the ones it’s trained on?
Bro did you record this video just after waking up or you got a new microphone 😭
i recorded it in the middle of night. couldn't sleep well
@@theAIsearch What got you to not sleep well? that sucks.
Watching this video while listening to AI-generated Lofi hip hop jazz in an ironic way, an experience as natural as industrial food
How to put our pictures in like our face ..I need to try this on ...like me as astronaut or me on statue of liberty...is it possible to use our images !?? to generate ??
This legitimately unnerved me
I can usually tell that there's something off about the eyes in ai generated faces but those images I dont see that. Woah!
Is this a reupload? I could swear I saw this video before. Or maybe just the images were the same...
14:24 _"Oh my goodness! Here we have six fingers on this hand."_
*"Hello. My name is Inigo Montoya. You killed my father. Prepare to die."*
The AI is confusing "dragon" in Komodo Dragon. It gives it fantasy dragon details but also has reference of a Komodo Dragon being a lizard.
all nice and fine so long as you want to hit the image generation slot machine, but kinda useless without being able to put in specific prompts