Dear Scott: You are the light in my day. Thank you for all the work, ideas, concepts, links that has been so helpful in my own work. As a graphic designer/illustrator, Midjourney has taken so many of my blinders off...and you have helped too.
I put in a photograph that I took to see what prompt it would suggest to get that photo, and then prompted Midjourney with the clip interrogator's suggestion. Really fun!
This is kind of funny to me, because image to text is literally how this whole image generation thing got started. Clip matches images with the correct text label and creates a text/ image imbedding, and algorithms like glide, Dalle, Midjourney, Stable diffusion, add diffusion and reverse the process to get text to image. The title to this video was a bit like seeing "Now available for the first time ever, the transcripts to books on tape!"
This has been super helpful for me. For one particular project on MJ, the artist whose style I wanted to re-create is not one MJ “knows.” I was able to come up with enough keywords and similar artists to pretty closely re-create the look I was going for!
Dude, I just did 30 of these. Got them all from Midjourney from other people's prompts. I feed the images to Google Collab, then I feed those Google Collab text into Stable Diffusion 1.5. Then I feed SD 1.5 the original text on Midjourney to compare. Google Collab did a better job at the prompt details than the original artist on Midjourney. It's like Midjourney preselects certain flavors and settings for you, while SD 1.5 you have to feed it as much information as possible to get the same level of detail like in Midjourney. The Google Collab text to imgs looked almost exactly the same as Midjourney than the original text on Midjourney. Crazy.
That bottom line of text is really good, I got a really good outcome with keeping it simple, where before like I was not getting anything very great. Vague prompt with more focus on the style/technics seems to pay off.
It is interesting to note that the current tools like DALL-E and Midjourney actually came out of THESE tools that take an image and describe it. Someone thought "Well if we can take an image and describe it, can we go the other way around?"
I prefer 'prompter' to artist. Using midjourney doesnt make me an artist. I feel only someones ego would make them think that. Obvs if actual artists are using MJ they are still artists
Yeah the way I see it, MidJourney/DALL•E/Stable Diffusion/etc. are the artists, as well as the canvas and tool/medium. The human providing the prompt is the director, producer and/or writer (analogous to the roles for creating a movie or show). If we adjust our language to be more accurate then we don’t need to feel guilty or judge each other. Directors, producers, and writers are very creative, respected and valuable roles in their own right.
This is great thanks for highlighting this. This is something I had asked if MJ was going to do, let us decrypt how the AI breaks down images. I've wanted to use something like this to decode some of my own past hand drawn work and see if I can recreate a style I love to draw in to speed up my workflow. I think this might also help buck the trend of people always using "trending on [such and such art platform] and let them tinker with things under the hood more specifically.
the other day I was talking with somebody and I was asking "I wonder if there is a tool that does it's best to try and workout how an image was created", well this is pretty much what I was looking for lol, awesome thanks Scott
No, it will not help. What I would do is use the same seed with a higher cfg and reroll. That should give you slight variations and do that until you are pleased with the image.
Haha, this had me choked ..... Marco Mazzori got properly pooped on. All that work perfecting his artistic ability and now anyone can produce by hitting some buttons and making out they are the artist, What a world we have created!
A great instruction, thanks. I just cannot manage to follow up where you drag your own work intothe Midjourney bot and obtain a link from that. That is at 2:40 minutes. Do I fail because I have no private channel or something?
right click on the image then select .. open the image in a new tab .. then drag the image from the new tab into a new tab .. you will see an "+" at the top of your browser drag the image into the "+" there you will see the cdn.discord... address
@@sedetweiler It does indeed give at least a detailed starting prompt for any image (non-AI) you upload. I'm using Colab, and I have to use only one of the "models" due to memory.
Amazing how fast this is all going, its like releasing a monster that will change art forever and I feel its going to devalue physical artists in the long term, if everybody could take a pill and become Usain Bolt, he then becomes lost in the crowd, its a scary thought to the long term full impact of this technology on art, one thing with the AI it sure does nail the lighting and that alone can make a poor image great to look at, this is all about to explode I feel, everybody can be John Sargent, make the fail-safe wow creation with a button giving the image life, and then paint it. The AI taps into the mind of artists connecting text/thoughts to copy their strokes/style, I love the technology I think its going to have a massive impact on art as we know it, scary.
I actually think the opposite will happen with those that can actually paint. As the market will probably become desensitized to digital art, those that can still fling a brush around with some skill have a real opportunity to have true value places on their skills. Since none of this is going away, everyone will need to adapt. This is akin to painters freaking out when cameras became a thing, or when film encountered digital cameras. All of the cheese just moved.
You have to use that as part of the story. What is she standing on, etc. Also, a portrait orientation will leave enough room for that pose. I would also try and avoid calling attention to the face with terms like smiling, beautiful, etc.
Looks like a great tool! I do get a problem though, when I insert my discord image link. An error comes out saying "NameError: name 'Image' is not defined" Tried with the standard cyberpunk image that is there as well, same result. Anyone having the same problem?
I had the issue as well but figured out the issue for me. Instead of immediately running the script, make sure you run the 'Check GPU' and 'Setup' tabs first! Hope this helps. :)
It would be an interesting tool for generating image metadata tags from an edited selection of terms for searches across thousands of images.. I love the idea of using it for research....it might be fun to find that other artist noted and feed it into the original generative text and see the results.
been looking for something like this, I found a few boorus that does it for anime style images, haven't tried it, but I think this model can be used to do the same.
Wow! really appreciated the detail and effort you put into this. I recently used Blue Willow to produce some incredible arts. If you could create a tutorial series about BW, that would be fantastic for the beginners like me
Omg I was just telling everyone how I wish I could figure out whose “style” or whatever, to use. I don’t know squat about this art terms or artists. This is exactly what I was looking for! You absolutely made my day. I was going to go to bed 2 hours ago but I’ve been so immersed in this I’ve been testing out a dozen images and found myself tweeking the words.
I'm confused, you can get the exact prompt they used almost 90% of the time by copying the command or prompt when viewing the image in the community gallery... in midjourney at least, all prompts used are available to everyone.
The big problem is: no matter how precise and accurate and detailed you make a description, MidJourney is not yet in the stage of converting that prompt in something that you imagined. It just forgets critical chunks of your prompt.
As all of these new products evolve, I am sure the tools we all use will also conform to them. It's going to happen as we have a lot of people with different goals all in the same game.
It doesn't forget something that it didn't interpret to begin with. Also, with each reroll interation, it tries to cycle to the next discription that wasn't focused on the first attempt. My complaint is that 90% of the non --test outputs are basically distorted garbage. Thank goodness for the new --test
@@SpiritTracker7 I'm not sure if it's just my impression, but --test is very weak at creating chimeras (hybrids) as compared to --v3. Anyway, I'm experimenting and exploring, within my limited spare time.
@@sedetweiler If we stop to think a moment, the algorithms are evolving at a frantic pace. And the big response from the users certainly plays a big role in the devs priorities... I'll say again: not that I'm complaining, things are fascinating as they are now.
this is true, but it is a place to start and learn some new artists we might not know. moreover, it is a lot of great ideas for prompts we might not have thought of.
Even if u used the same prompt twice you can't get the same Image....in the future art will be too one do.entional..u have to have lots of different styles to be an artist. And those who purchase can't just purchase one it has to be something that changes it's form into many.
When these AI generators came out, everyone assumed that artists would lose their jobs. But some people invented careers like Prompt Engineering and prompt marketplaces where you sell your ability to write a prompt. Now, prompt engineering can be automated by reverse-prompt It's fascinating that we had this tool that could do something for free.. and instead of everyone creating and sharing freely, we invented a different specialization and marketplace. If someone invented a 3D-printer for food, would we invent a new copyright to stop anyone printing food for free?
All of those artists , movements etc. are all hard coded in. The CLIP Interrogator AI only checks match percentage against them. That hard-coded list is not exhaustive. Which is why it doesn't get it always right. Someone needs to scrape all the discord and galleries and make a bigger list with all the top ones.
@@sedetweiler No, not really. The model this is using is called CLIP. It has different versions ( that was the selection you had in the Collab ) but the one Stable Diffusion uses is currently fixed ( it's the one they tell you to choose for SD in the colab ). The tokens ( think words ) supported by this model are the only ones understood by Stable Diffusion. But that's a long list. Around 30k. It's not possible to check similarity to each of those words in the colab, so the colab has a fixed hard coded list of artists etc. it checks against. That list is small and not really updated regularly. That's what I'm talking about. The names and words in the Collab.
The interface has changed, and I cannot figure out how to use an image URL.
Dear Scott: You are the light in my day. Thank you for all the work, ideas, concepts, links that has been so helpful in my own work. As a graphic designer/illustrator, Midjourney has taken so many of my blinders off...and you have helped too.
Thank you! Your comment made my day! 🥂
I second that, thanks so much.
Thank you!
I put in a photograph that I took to see what prompt it would suggest to get that photo, and then prompted Midjourney with the clip interrogator's suggestion. Really fun!
It's a pretty fun way to get some new ideas!
Great way to learn what prompts the AI can look for. Great idea and tool!
It really is!
This is kind of funny to me, because image to text is literally how this whole image generation thing got started. Clip matches images with the correct text label and creates a text/ image imbedding, and algorithms like glide, Dalle, Midjourney, Stable diffusion, add diffusion and reverse the process to get text to image. The title to this video was a bit like seeing "Now available for the first time ever, the transcripts to books on tape!"
Yeah...it's more like..now available to the PUBLIC! They've always been able to parse this data internally. At least I think so.
This has been super helpful for me. For one particular project on MJ, the artist whose style I wanted to re-create is not one MJ “knows.” I was able to come up with enough keywords and similar artists to pretty closely re-create the look I was going for!
Glad it was helpful! That is a pretty great use of this tool!
Dude, I just did 30 of these. Got them all from Midjourney from other people's prompts. I feed the images to Google Collab, then I feed those Google Collab text into Stable Diffusion 1.5. Then I feed SD 1.5 the original text on Midjourney to compare. Google Collab did a better job at the prompt details than the original artist on Midjourney. It's like Midjourney preselects certain flavors and settings for you, while SD 1.5 you have to feed it as much information as possible to get the same level of detail like in Midjourney. The Google Collab text to imgs looked almost exactly the same as Midjourney than the original text on Midjourney. Crazy.
It seems that when I follow the instructions with Discord I run into an error that can't resolve. Is this something others are experiencing?
That bottom line of text is really good, I got a really good outcome with keeping it simple, where before like I was not getting anything very great. Vague prompt with more focus on the style/technics seems to pay off.
It seems that the bot just does better when it's given leeway. I don't use really long prompts very often.
It is interesting to note that the current tools like DALL-E and Midjourney actually came out of THESE tools that take an image and describe it. Someone thought "Well if we can take an image and describe it, can we go the other way around?"
Yup! I have an FYI video coming that tells that story. You are ahead of the curve!
Thank you so much for holding us on ! Definetily deserves a like (and some support on Patreon when i'll finish my studies and get a job ^^)
I will keep them coming and you go finish school! ;-)
And Scott, as always, bringing super new wonders for us, humble mortals....
I am just doin' my best :-)
I prefer 'prompter' to artist. Using midjourney doesnt make me an artist. I feel only someones ego would make them think that. Obvs if actual artists are using MJ they are still artists
Not a bad way to put it, actually.
Yeah the way I see it, MidJourney/DALL•E/Stable Diffusion/etc. are the artists, as well as the canvas and tool/medium. The human providing the prompt is the director, producer and/or writer (analogous to the roles for creating a movie or show). If we adjust our language to be more accurate then we don’t need to feel guilty or judge each other. Directors, producers, and writers are very creative, respected and valuable roles in their own right.
This is a nice way to put it.
What about wizards? Are prompts not unlike spells, making the impossible possible?
@@MP-wp8cu not remotely, no. It is so possible that my 6 y.old son can do it...
Does this work on regular images, one that was not created by text to image AI?
Yes! Absolutely!
@@sedetweiler Sweet! That makes it an invaluable tool for prompt crafting!
Total agree! It has a lot of hidden potential that way.
This is great thanks for highlighting this. This is something I had asked if MJ was going to do, let us decrypt how the AI breaks down images. I've wanted to use something like this to decode some of my own past hand drawn work and see if I can recreate a style I love to draw in to speed up my workflow. I think this might also help buck the trend of people always using "trending on [such and such art platform] and let them tinker with things under the hood more specifically.
I see so many of the same messy prompts and most people have no clue what they all mean. you are so right!
the other day I was talking with somebody and I was asking "I wonder if there is a tool that does it's best to try and workout how an image was created", well this is pretty much what I was looking for lol, awesome thanks Scott
You are welcome! It sure is a fun one to experiment with!
ive been looking for a free imige to prompt thing! thank you !
Curious why you didn’t use the checkbox for Stable Diffusion.
I was probably side tracked. I noticed that after I uploaded it. :-)
@@sedetweiler Ha! It happens. Thank you for the reply. Love and appreciate your content. Keep up the great work!
Thank you!
I like use > "I make Promptism", the isms of art
Hi Scott,
Love your videos.
Are you aware of any methods of reverse image promoting a large batch of local files and spitting them out into a CSV?
There are some extensions that do that, I believe.
This is great, I wanted a tool similar to this. Thank you.
This was what I was looking for, but it still doesn't work the way I would like it to.
It's pretty flexible this way.
Scott, I find the woman's heads to be mis shapened in SD. Would raising the STEPS or SCALE help?
Unfortunately, no. It's a matter of the model learning, but you can also use that same seed and it should create close alternatives.
No, it will not help. What I would do is use the same seed with a higher cfg and reroll. That should give you slight variations and do that until you are pleased with the image.
Is there an AI that can scan the image and generate its respective LUT? That would be really interesting if there is.
Not that I know of. I am sure color grading isn't hard to adapt. I know you can do it in Photoshop with a series of adjustment layers.
Wish it would store engine, settings and prompt in EXIF information.
I have suggested the same thing! I am sure someone will add that at some point. I use Google Keep for all of my favorite prompts right now.
Is there an updated video for this site, my current options don't say interrogate ?
I figured it out, thank youu so much.
Can you please share the link
Haha, this had me choked ..... Marco Mazzori got properly pooped on. All that work perfecting his artistic ability and now anyone can produce by hitting some buttons and making out they are the artist, What a world we have created!
A great instruction, thanks. I just cannot manage to follow up where you drag your own work intothe Midjourney bot and obtain a link from that. That is at 2:40 minutes. Do I fail because I have no private channel or something?
I just created a dm to the bot. You could literally throw it into any discord channel.
right click on the image then select .. open the image in a new tab .. then drag the image from the new tab into a new tab .. you will see an "+" at the top of your browser drag the image into the "+" there you will see the cdn.discord... address
@@sedetweiler Thank you for the suggestion. It sadly does not work for me. but I will keep trying. You might have different settings.....
Awesome, thanks for sharing this!
Glad you enjoyed it!
when you say that you photo shop all over the top of your artwork what does that mean exactly? Great videos by the way.
I like to combine it with other photos and images
Nice video! How can I have that Midjourney Bot in my discord?
they have instructions on their discord. It wasn't difficult.
Can I use stable diffusion for this method? What would I need to do?
Is this only for MJ/SD? What if you dragged a non AI image? would it tell you what it thinks it would be?
Yup, it would work.
@@sedetweiler It does indeed give at least a detailed starting prompt for any image (non-AI) you upload. I'm using Colab, and I have to use only one of the "models" due to memory.
It should work well.
thanks for video . I wonder how much cost we need for start exploring those tools including Private mode gallery and image editor for fix some error ?
Private mode is $20 more, so $50 a month total currently, but I am sure that will change.
thank you very much! it is very helpful!
Glad it helped!
Does not work anymore. 403 error on 3rd step
super .. thx !!
This is cool. Thanks
Sure!
Stable-diffusion CLIP interrogator: "a man sitting in a chair with his legs crossed and a beard on his head..." close but no cigar.
OMGODMAN I love you!
😂
Scott seriously are you half 🤖 🤣!? Love all your explanations on your clips ❤️ 🙏🏼
I wish! lol!
Amazing how fast this is all going, its like releasing a monster that will change art forever and I feel its going to devalue physical artists in the long term, if everybody could take a pill and become Usain Bolt, he then becomes lost in the crowd, its a scary thought to the long term full impact of this technology on art, one thing with the AI it sure does nail the lighting and that alone can make a poor image great to look at, this is all about to explode I feel, everybody can be John Sargent, make the fail-safe wow creation with a button giving the image life, and then paint it. The AI taps into the mind of artists connecting text/thoughts to copy their strokes/style, I love the technology I think its going to have a massive impact on art as we know it, scary.
I actually think the opposite will happen with those that can actually paint. As the market will probably become desensitized to digital art, those that can still fling a brush around with some skill have a real opportunity to have true value places on their skills. Since none of this is going away, everyone will need to adapt. This is akin to painters freaking out when cameras became a thing, or when film encountered digital cameras. All of the cheese just moved.
douse anyone know how to have this tool run on my own pc, since the server reach it limit sometime
You need to have an Nvidia 3080 with 10gb of ram. That's the lowest I would want to use.
Would this work for Dall-e 2 as well?
I don't see why not. But, keep in mind that Dalle was probably trained on mostly stock photos, so the results are going to look more sterile (imho).
Runtime Error :/ Something something can't allocate
Running in Chrome?
I'm still trying to uncover the prompt that gives an entire body, no hair or legs cut off. Anyone?
You have to use that as part of the story. What is she standing on, etc. Also, a portrait orientation will leave enough room for that pose. I would also try and avoid calling attention to the face with terms like smiling, beautiful, etc.
Looks like a great tool! I do get a problem though, when I insert my discord image link. An error comes out saying "NameError: name 'Image' is not defined"
Tried with the standard cyberpunk image that is there as well, same result. Anyone having the same problem?
I had the issue as well but figured out the issue for me. Instead of immediately running the script, make sure you run the 'Check GPU' and 'Setup' tabs first! Hope this helps. :)
@@estabanpls It worked! Thanks!
this is a great tip! thank you!
👍🏻
🥂
It would be an interesting tool for generating image metadata tags from an edited selection of terms for searches across thousands of images..
I love the idea of using it for research....it might be fun to find that other artist noted and feed it into the original generative text and see the results.
there is a lot to be learned from this for sure, and I think you are on the right track!
been looking for something like this, I found a few boorus that does it for anime style images, haven't tried it, but I think this model can be used to do the same.
🧐🧐 Blows my mind how this is even possible....redonculous.....
It's guess it's just using the bot to try and do the math backwards, in simplest terms.
I totally had to check out your channel with a name like that! That was pretty awesome!
@@sedetweiler haha...yeah i've been slacking on uploading...but have a ton of video I need to get out 🐓🐓
Do it! ;-)
Wow! really appreciated the detail and effort you put into this. I recently used Blue Willow to produce some incredible arts. If you could create a tutorial series about BW, that would be fantastic for the beginners like me
I'm thrilled to hear that you had success with Blue Willow! I think a tutorial series would be a great way to help beginners get started with the tool
Getting error NameError Traceback (most recent call last)
Omg I was just telling everyone how I wish I could figure out whose “style” or whatever, to use. I don’t know squat about this art terms or artists. This is exactly what I was looking for! You absolutely made my day. I was going to go to bed 2 hours ago but I’ve been so immersed in this I’ve been testing out a dozen images and found myself tweeking the words.
me just wanting to know the model art
I'm confused, you can get the exact prompt they used almost 90% of the time by copying the command or prompt when viewing the image in the community gallery... in midjourney at least, all prompts used are available to everyone.
Unless you pay for private mode, like I do. Or, maybe it's art you enjoy that wasn't made by AI. 😉
This works for every image, not only those made with MJ.
The big problem is: no matter how precise and accurate and detailed you make a description, MidJourney is not yet in the stage of converting that prompt in something that you imagined. It just forgets critical chunks of your prompt.
As all of these new products evolve, I am sure the tools we all use will also conform to them. It's going to happen as we have a lot of people with different goals all in the same game.
It doesn't forget something that it didn't interpret to begin with. Also, with each reroll interation, it tries to cycle to the next discription that wasn't focused on the first attempt. My complaint is that 90% of the non --test outputs are basically distorted garbage. Thank goodness for the new --test
@@SpiritTracker7 I'm not sure if it's just my impression, but --test is very weak at creating chimeras (hybrids) as compared to --v3. Anyway, I'm experimenting and exploring, within my limited spare time.
@@sedetweiler If we stop to think a moment, the algorithms are evolving at a frantic pace. And the big response from the users certainly plays a big role in the devs priorities... I'll say again: not that I'm complaining, things are fascinating as they are now.
"It may guess right, it may guess wrong", I can do that 🙄
this is true, but it is a place to start and learn some new artists we might not know. moreover, it is a lot of great ideas for prompts we might not have thought of.
Even if u used the same prompt twice you can't get the same Image....in the future art will be too one do.entional..u have to have lots of different styles to be an artist. And those who purchase can't just purchase one it has to be something that changes it's form into many.
Coach Germaine 😬😡😬
well said. ;-)
When these AI generators came out, everyone assumed that artists would lose their jobs. But some people invented careers like Prompt Engineering and prompt marketplaces where you sell your ability to write a prompt. Now, prompt engineering can be automated by reverse-prompt
It's fascinating that we had this tool that could do something for free.. and instead of everyone creating and sharing freely, we invented a different specialization and marketplace. If someone invented a 3D-printer for food, would we invent a new copyright to stop anyone printing food for free?
All of those artists , movements etc. are all hard coded in. The CLIP Interrogator AI only checks match percentage against them. That hard-coded list is not exhaustive. Which is why it doesn't get it always right.
Someone needs to scrape all the discord and galleries and make a bigger list with all the top ones.
I am sure this is always being retrained. I know that takes as much as 3 weeks with the larger models, but it is only a matter of time.
@@sedetweiler No, not really. The model this is using is called CLIP. It has different versions ( that was the selection you had in the Collab ) but the one Stable Diffusion uses is currently fixed ( it's the one they tell you to choose for SD in the colab ).
The tokens ( think words ) supported by this model are the only ones understood by Stable Diffusion. But that's a long list. Around 30k.
It's not possible to check similarity to each of those words in the colab, so the colab has a fixed hard coded list of artists etc. it checks against.
That list is small and not really updated regularly. That's what I'm talking about. The names and words in the Collab.
not funtion!