Leonardo AI - Create Consistent Characters
ฝัง
- เผยแพร่เมื่อ 1 เม.ย. 2023
- In this video, I will show you how to generate consistent looking human characters in stable diffusion using LeonardoAI. Stable diffusion is a technique that allows you to create realistic and diverse images from a low-dimensional latent space. LeonardoAI is a platform that lets you train and use stable diffusion models for various domains, such as faces, animals, landscapes, etc.
This method works with any stable diffusion model on any platform, including Automatic 11111, InvokeAI, Midjourney AI, and others. You can use it to create characters for games, comics, animations, or any other project that requires human faces.
Links:
LeonardoAI: leonardo.ai/
Reddit Post: tinyurl.com/2jrtsny9
☕ Buy me a Coffee: ko-fi.com/promptengineering
Keywords: stable diffusion, LeonardoAI, human characters, names as anchors - วิทยาศาสตร์และเทคโนโลยี
If you liked the video, check out this video on how to create your own AI avatar: th-cam.com/video/V2efVSXSlqc/w-d-xo.html
yes
Characters doesnt look real at all.
I am happy that i came across your channel! I want to use Leonardo AI more extensively from now on and your videos are really really good and well explained.
Dude, you just decoded the Matrix! thank you.
You are welcome Dr B :)
So the key point and step 1 is to use unique name, right? Step 2 is collect images that you like for dataset. Step 3 is train our own model using Leonardo option. Step 4 is create unlimited images with different outfit and different poses , but with the same face(character) using your own model. Am I right?
Pretty cool. I'm trying to adapt it to characters I've already created using SEEDS and other parameters. The truth is that the result is much more consistent than just using a SEED without more. EXCELLENT!
Glad you found it useful!
just starting out on my AI art journey and this is a really helpful video - many thanks for sharing!
I'm similing, great video, I did not know about the use of names in prompts
Its a great trick :)
i found this video interesting, informative and and fun to watch.
thanks for the upload
That's a good idea. It has to create some sort of value for something even if it doesn't know what it is and it's going to generate the same value every time. It's just a matter of finding which name creates the person you want.
Exactly.
Simple and easy tutorial, thanks 👌
get up and dance along, Keep up the great work, guys!
Thanks so much for sharing! Was looking for this :) especially in Leonardo AI.
Question: if you create a person with a certain look first, are you able to tie that look to a name and generate different scenes from there? Because I want my characters to look a certain way, not just determined by what comes out from the machine when I use a name. Cheers! 🙏
There are a couple of options. You can use that image as part of the prompt with a fixed seed. This works in MidJourney (I have a video coming up on that next week). I haven't tested that in Leonardo ai so not sure if it's going to work here. The second approach will be to train a dreambooth model of your person. If you are using an ai generated image, then you probably will not have enough images of that character to train a model. So I would recommend to first use img2img (with high weight) to first generate a dataset. Hope this helps.
@@engineerprompt thanks, will try first option first! If not working, I will explore dreambooth. Appreciate it! Keep up the good work :)
Brilliant! Got to try it out.
This makes me wonder, would it be possible to have image to text guess a persons name from a photo or series of photos, and then use the same name with text to image to generate more images of that person without ever training a model.
Pretty amazing, thanks a lot for that
Do you think there's a hack to do the same thing using a custom picture ?
Pretty cool your trip . It generate pretty cool portraits 😊
Great tutorial, thank you.😊
Great video! Thanks!
Super helpful, thank you!
Thank you 😊
This is brilliant.
If you mention a website you should include a link in the description. What is the random name generator web page?
Nice tip mate, thx!
You didn't address the big question: does the unique 'name' of the person carry over to other sessions, giving same results? I guess not, but it should if the AI companies got their act together. One should be able to count on there being consistency from session to session in some manner.
Thanks for this Tutorial
What about creature characters?
Could you please make video on "Character Consistency SOLVED in Stable Diffusion" since i could not find it on youtube?
Hi do you need this app if you already have runway?thanks
How owuld you prompt to create specific pose and hand position of your human model? I need to create an athlete that is holding his hands up in front of his chest.
Very helpful. What if I have want to use an existing picture? For example, I have a headshot photo, and I want to generate a full-body image, and then use that as a consistent character (able to change clothing, poses, etc.)? I want this to look like a real person, not an avatar.
this is what I spent most of my day trying to figure out, watching one video after another and none of the different tools i've tried have created an image that looks like the person I upload. It seems like it should be the most simple thing for AI to do, right? Copy this image, give the person new clothes, a different pose. nope.
That's what the custom models creation are meant for, you can compound and add seed input when you have a look you like. Or use image to image
@@lumideo- What is ‘Seed input’? ( Sorry I’m new to Ai image generators.. lol)
How do i generate full body image ?
Thank you so much Bro
Hi. I would like to add one more additional rule after naming a character. You need to delete previous generated images which do not have the desired face or look because the AI will use previous images as references.
Interesting insight, that makes sense
It works for me. I hope it works for you too 😁
Are you trying to say that AI is being trained on the images we create? I doubt that.
Wow, did not know this. Always wanted a way to 'tell' the AI that X, Y look good, but Z looks bad. I'll try it!
Thank you soooooooooooo much!!!!
Thanks as always. Is it better to fix the seed to keep the routine character or Laura? or midjourney? Leonardo AI?
Seed plays a critical role. If you can train a LoRA, that's the best option.
I noticed They had almost similar faces even though they had a different name.
Is there a way to name the character if we got the face first and then call it back to get the same facial structure?
For that you will need to look at training your own models. Look into dreambooth and loras
@@engineerprompt I know that stuff, what I mean is that if we encounter some character we generate it from our prompt and "name" that character ourselves so we can call it back without Lora or some embedding.
or it's just not possible for now?
@@MrSongib As far as I know, that's not possible at the moment. In this case the "Name" is used in the latent space to derive the creation of the image in the first place. But giving a name after creation of a character is not really going to help in the next image creation step.
@@engineerprompt Yeah sadly.
But if I used your method here seems good and can recall the face pretty well. so I think for now is that put some random names until you like the character that you get and recall them again in the future project. it's working pretty well. thanks.
Does this work for objects?
Can you do this with an image you want imported?
Thank you, but I tried the free version and I don't get the pictures I desire , they give me cartoon pictures which I don't need. Is the pictures you generated belong to the paid version?
Is it possible to use this method with a custom character from a user's photography and to use another at the same time another image to define the posture of the character?
In order to do that, have a look at pose transfer (th-cam.com/video/2vICGO12ji8/w-d-xo.html)
hi, do you have any experienced with image generation for fashion? like maybe outfit of the day but with the same character? thank you
Yes, you can check out playground which has that ability.
Impressive video! I do character design and Stylar AI has been a recent discovery. This AI tool is fantastic for designers, offering control over AI models and image results. It's still developing and somewhat buggy, but I highly suggest trying it.
I was using random name, it was giving similar face but when i made changes in promt like "clean shaved" then it start giving me totally different face,and then i Lost my previous character...i think in very beginning we should give promt in complete details otherwise character will be change
can you copy paste the nagative promops?
Can you make a video on how to stop Leonardo from cutting the head off of every render? I have my own fine tuned model made up of 11 768x768 close up photos. When I try to use them, and start with full body, eye level, zoomed out etc, and for neg I put zoomed in, close up etc. I am spending all of my time just trying to get a full head, and have no time for creativity.
Probably use something like ultra-wide angle in the prompt. That might help. I haven't looked at Leonardo in a while but will be looking at it soon as they have added some great features.
Really enjoyed your video! In my character design work, I've been exploring Stylar AI, an AI-based design tool that's very designer-friendly. It provides great control over AI models and image outputs. It's in its early stage and a bit buggy, but trying it out could be beneficial.
Is it free?
Iv created 3 cartoon character illustrations but I can't recreate them in different scenarios without the character completely changing even if the character has been named. Any help or ideas guys??
Your best bet will be to train a dreambooth or lora for that case.
@@engineerprompt ok thanks for replying. Now all I need 2 do is learn wotever that was u said haha
@@sekiroskedro 🙂
Why not photorealistic instead of deliberate 1.1 model?
Do you know a trick/the best way how to get the same person with the same outfit/style but in a different position? Like waving, running, looking down etc.
Use either dreambooth or controlnet
@@engineerprompt Okay. Shall i use the unique name for the person in the following prompts as well? I am not very good with the prompt design...
Did you try this by just putting in random characters and numbers? maybe theres a benefit for making it x tokens long. I'm sure it doesn't have to be an actual name
From my experiments, very unique name does have an impact.
2:34 can i get the promtes for fixing emma watson face i could not even after trying for 2 days adding every negative prompt to fix this but i couldnot
Hey make a tool for 3d modeling text to generate a 3d model using AI
How can I keep a trained real person ‘s face dataset or image to image looking in consistency, I found face is not real always kept the same as trained photos, your hack seemed for virtual persons only , not for real general person
Can we generate our own images. By providing some reference images?
Yes, so the initial image can be yours and then assign a name it to. It's probably not going to be too stable though. I would recommend to look at something like dreambooth for that purpose
im similing right now
Sorry as a Neophyte to AI and SD 1.5, disagree There are several ways to generate different images not only with the same face but even with different styles and no need for any LORAs. One of the major keys items we have found is the SEED, not always need the seed, but with it one can get more consistency. Also, no a name is not needed. But what is being done by using a name is narrowing the AI sampling set to a sample that already has a consistent facial image and likeness. Such as when you are using Emma's name. The AI sample is pulling from Emma W samples. We wish that you could interrupt generation and save the image at a particular percentage because not only do you see the sample sets, but sometime the image is even just as good as the end results.
In our short time we have created characters with same facial image without training in our journey to learning SD. Several we have added to the NSFw galleries. One of the lates is The AI model "Staci" and now the name "Staci" was not part of the original prompt.
Cut me off! Thx rosey!
I’m gonna attempt to make a Noghtmare on Elm street prequel using video A.I……. I already wrote a insane bad ass script……..Robert England identical voice and face…….wish me good luck guys!
Good luck👍
keep your attempt, things getting easier
@@BenPanna That's true. A little more patience. Bigger things are on the horizon. I also got this feeling Holiwood is planning to block the next gen AI tech for films and animation.
Which is better? Playground AI or Leonardo AI?
Both are good, if you want more models, go with LeonardoAI. But Playground is great if you want to edit images with text like instruct2pix. You can select an area of the image to edit. That's a pretty neat feature.
@@engineerprompt have you played around with Playgrounds infinite canvas? That's a really cool feature, works well. Haven't tried Leonardo but Playground has added a new model called Playground which is pretty nice
I tried this with Indonesian common names and it's not working, the faces are very differrent.
Then i tried with celebrity names, and voila, it generatese similar face of the celebrity. I think it's not working on "random" names, it has to be someone whose face is recognizable enogh in the internet, at least it's true with Indonesian names.
8:22 Liga Reddit del truco para generar imágenes realistas de personas, estilo foto real de frente y a nivel del busto, para aplicarlo en Leonardo IA . El chiste también es generar nombres únicos aleatorios desde una página 📃 generadora de nombres, según el país o la época.
please comment the negative prompt which you have used
Each model on Leonardoai has a separate negative prompt that it will use.
Did a quick check on FB. It appears that the names you typed in coincide with the people on FB. This is not a trick, the program just uses the real people it finds linked to the names on FB to generate the images. 😒
Interesting, the model doesn't have access to internet after training. My suspension is that it probably learns average look for each region
Bro can you tell me how to create accurate picture of celebrities?
Check this out th-cam.com/video/hazFoGUKE8Q/w-d-xo.html
confusing. Why are you inventing a random name and using it? Did not see why you were doing that.
so complicated...
Dude we can't hear you
You talk too fast and your words are not enunciated great. It's hard to understand some of what you are saying. Just FYI
If you're gonna shill for a website, you should make it obvious your video is an add.
it's quite hard to insit on one AI tool, too many review that's hard to choose
I am not getting paid by them (I wish I would, will help with the bills). All these different platforms have their own strengths and weaknesses. Just need to pick the tool which is best for the task at hand.
For Christ sake... can you higher your audio...
Sounds good to me