This New AI Tool Does Consistent Characters (For Real)
ฝัง
- เผยแพร่เมื่อ 3 ม.ค. 2024
- Consistent AI characters have long been the holy grail of AI Image generation, and I'm happy to say: It's finally here.
🔥 Discover the ultimate guide to creating consistent characters in AI imagery using Scenario! Join us as we delve into this groundbreaking AI image generator.
👉 Learn through a comprehensive guide on crafting custom characters with varied aesthetic styles and discover how to optimize Scenario’s features for character consistency. \
FREE PDF WORKSHEET and TEMPLATE CHARACTERS:
theoreticallymedia.gumroad.co...
LINKS:
Scenario: app.scenario.com/?ref=theoret...
👉 Stay Connected:
Twitter: / theomediaai
Join the Discord: / discord
🙏 Support the Channel:If you'd like, please join the Patreon Here: patreon.com/TheoreticallyMedia
Like, Comment, and Subscribe if you enjoyed the video.
Hit the bell icon to stay updated on our latest uploads.
Chapters - วิทยาศาสตร์และเทคโนโลยี
Download the FREE PDF here: theoreticallymedia.gumroad.com/l/characters
Scenario: app.scenario.com/?ref=theoreticallymedia
Yo! 👋
@@LouisGedo THERE IT IS!!!
Great info. Cool basses on the wall.
@@TheoreticallyMediawe can easy Made manga with this
Great guide Tim!
Thank you :)!
Why would you choose to use a famous character that the model already knows? That makes this demonstration of it's capabilities fairly pointless
I would say that you can almost always find a reference somewhere that will help the computer approximate what you are looking for. After all, it’s what our own brains do - you cannot imagine something that is nothing like anything you’ve ever seen before. Everything you picture has roots in something you have seen
totally agree, we only dream of things we have seen before or thing we have given a thought while we were woke. BUT this is AI, not human, do it should not think like humans. I agree you can always findvs similar model but similar is not the real thing.
it still useful. AI like any creator requires references. In order for it to be consistent you need to provide it with the character from different angles and perspectives with the same clothing etc. Same principle for a 2d concept art to be translated to a 3d model.
Heck YES this is exciting- I can't wait to try it out! Thankyou, another amazing tut as usual :D
Happy New year! thanks for all your videos!
Happy New Year to you as well!! And Thank you!!
"What a difference a day makes..." so the song goes, but 7 days! I really missed all your AI updates. This one is critical to my current project. Thanks Tim for your pointers.
Excellent! You might want to check out the follow up video (next one) as that has some tips and updates from Scenario!
It’s the Upscale video on Magnific, but the Scenario stuff is toward the end of the vid!
Really wonderful walk through!Thanks Tim, you are my favorite. ;)
Thank you so much! And really appreciate seeing you in the comments all the time!! Really means a lot!!
Incredible video, Tim! Watching you with Scenario is like seeing a magician pull a rabbit out of a high-tech hat. AI art's journey from "abstract blob" to "hey, that's a person!" is nothing short of miraculous. And here's a funnier AI joke for you: Why did the smartphone go to therapy? Because it lost its contacts and couldn't connect anymore! Thanks for the tech-tacular insights!
Great content, as always, Tim!
Thank you so much! Looking to revisit Scenario this month! They’ve added a lot of cool new stuff!
Wow, this looks exactly what I was looking for, thanks for informing me!
Excellent! So happy to hear that!
Fantastic, very excited about the projects I could use this in
100% same!
Thanks, Tim. Very informative as always. I just forwarded this to a friend wanting this exact type of thing for a board game. 👍
This is indeed a great platform for producing fairly stable characters. Been using it to generate scenes for generative video ai production. Takes a bit of work but you can get some pretty amazing results.
I was a little wary of introducing ControlNet, as that can get a bit wonky- but I figured as long as I mentioned it wasn’t 100%, it would be fine. I mean, you CAN get some amazing things out of it- but like you said: it takes some work!
Good stuff. What platform do you like to use with this to generate videos?
nice sharing and that is something we all want, consistent characters and being able to create a video with the character
Going to dig a little deeper into Scenario soon! This thing has a lot of tricks up its sleeve!
if you really wanted it you would develop drawing skills.
Hire a damn artist
Why? 🤔 I'm done supporting artist both on Patreon (who don't upload consistently whatsoever & use every excuse in the book as to why) and through commissions which they constantly extend the deadline while still overcharging. Get with the times and stfu. @_Cuertzia
@@extraspicymochi what you said has NOTHING to do with the concept of hiring an artist.
Plumbers don't show up or half-do jobs - but are you done with all plumbers and plumbing?
Mechanics half-do their jobs and overcharge. Are you done with getting your car fixed?
Clients are unspecific, drag their feet, pay late, and expand scope after the agreement. Artists CAN be "done" with clients... I suppose. But it doesn't make it wise. It doesn't make the problem universal. Some plumbers, mechanics and clients are really good and behave well and professionally.
Some artists do too.
But show off your anger, hate, condescension and bigotry somewhere else.
That some artists take longer to make things then they said on the outset is NOT an excuse to use ai.
If a house takes longer to build than expected - that's reality.
If a car takes more work to fix than initially expected - that's reality.
If it takes longer to approve a piece of art than expected - things happen. Sometimes it is unavoidable.
You're a world class jerk.
Hire an artist became a diatribe against patreon? Fine. Don't follow people on patreon. IRRELEVANT.
Do what you JUST did with some other group. Take a handful of black criminals - and start demonizing ALL black people over the actions of those specific examples. Say "you're done with black people."
Do it. Or admit your standard was unfair, hateful, bigoted and stupid.
Incredible. Thank you for the breakdown.
1000%! I have some more Scenario stuff coming your way soon as well!
Consistent characters, and correct looking hands, are the two biggest challenges.
Using the same seed and somewhat of the same description does help keep a character looking the same (as long as every single generation parameter is the same)... But getting proper hands is VERY difficult, only 1/10 images I generate have decent looking hands. I'm using epiCRealism and have tested many other models, seems like they all struggle a lot.
I did see news about a model that addresses the hands/fingers issue. It's basically faceswapping for hands-- but I haven't gotten the chance to play with it yet. So, we're almost there!
(Although I'm sure that v.1 will still have issues!)
Question - and you may not know and I apologize if this is beyond the scope of what you care to discuss - would it be possible to get hands right if you had a model that was trained on hands? Or would it still mess it up because the detail is too fine?
Having a very well trained Lora can improve hands. But it may take many attempts, a lot of experimentation with various descriptive words and images. I was going to make a few hand Loras but actually found it very difficult to find enough good images of hands. When you search a database for "hands" you end up with artistic images of hands in odd positions. When you look through images of portraits, you end up with very few good images of hands. Finding images of hands which are suitable to serve the sole purpose of correcting hands in AI Image Gen seems to be quite difficult, and this actually might even be the root of the issue for full models like EpiCRealism too.@@TheVoiceofTheProphetElizer
Thank you for this video, it's incredible and for the files too! I don't intend to use this for anything in particular but I'm interested in what this tool can do
Give it a shot for sure, I mean the free plan is pretty generous, so you can explore the other things it can do! And if you run across something cool, please let me know!
Thanks for the video, you helped me learn how to create characters based on 3D models. I used this survey in Leonardo AI and got a pretty good result.
excellent to hear!!
Yes, this. I have wanted consistent characters for so long.
Right with you! There have been so many fake out versions, where things worked (kinda), but ultimately didn’t.
Here. We have it. Finally.
(I mean, there’s still all the usual AI weirdness, but still: one big problem down!)
Thanks for that. I was looking exactly for this.
Congratulations on approaching 100,000 subscribers! I hope it happens soon. :)
Thanks for your effort. This is the first image generation to close in on what I actually want. Great stuff.
Agreed!! I'm really excited to see how this one develops! It's funny that its actual purpose is for game assets, but goes to show: Tool Bashing is a wonderful thing!
I had this in beta test since the in April last year and I cant promise you, it's really great!
Great video. Excited to try this out.
Does it work, have you tried?
Great stuff. And consistent characters in images inevitably leads to consistent characters in videos since we can do image to video. Loving it.
yup!! Midjourney is right on the cusp of introducing video as well...they claim to have a different take on it-- so, I'm REALLY interested to see where that's going!
Thanks! great tutorial video. I'm excited to try it out
Thanks so much! It’s really pretty good! Let me know what you think when you get to playing with it!
Your videos are so helpful and well presented. Thanks.
Thank you so much!
Have you had any issues with Scenario's model maker cutting off tops/sides/etc. of images you upload? Everything seemed to work flawlessly until I went to make the character model... but it won't let me scale the image when I try to adjust the crop.
Not too shabby & certainly an easier alternative to learning all the ComfyUI nodes I had to tinker with to achieve the same results. That said, I'm sticking with ComfyUI as it gives me far more control over everything. Thanks for the info & rock on!
Glad to hear! And yeah-- this is a great alternative for those who don't have the GPU horsepower or the time to dig into Comfy. I mean, I like Comfy as well-- and while I'd love to do a video on it, I just can't crack a way to simplify it for a wider audience.
Kind of my underlying mantra for the channel is to make all these creative tools accessible to everyone-- Comfy, while pretty awesome-- gets fairly complex really quickly.
Although-- now you have me thinking about a tutorial that's something along the lines of "If you're ready to get Advanced with AI Art"-- That might be something interesting for the intermediate crowd who are ready to stretch out of Midjourney?
Yeah, Comfy's a lot to wrap one's head around (mine's still spinning, but I'm learning a lot) & it's terrific that there are so many fantastic tools & toys available that take some of the guess work out of it. Accessibility is a wonderful thing! As for intermediate tutorials, why not, eh? If you're already getting their feet planted on solid ground with the basics, why not show them where they can go with it? :) Keep up the great work!
As long as the story and characters are original this is amazing
The reason the AI used elf-ears is that your character reminds me a lot of a comic book character called Cendre, created by the artist Orpheelin many years ago. The comic book is called Northern.
The backwards facing sword is actually a well-used killing technique where the hand is drawn forwards and down fully across the body surprising your opponent in front of you. It was well used in swordplay in Asia and Europe.
AND! It looks super cool!! Haha-- that's great to know though! So, technically-- I guess she wasn't wrong!
this is awesome! Thank you!
Glad to hear! I do have some additional tips and tricks at the back end of this video as well: th-cam.com/video/unqRqIDyHuI/w-d-xo.html
Looking to do a big deep(er) dive into Scenario coming up really soon as well!
Really good video, thanks 👍
I can't wait to have a consistent character option, storytelling simply doesn't work without it. Thanks for the video, it's definitely a step forward in the right direction
Indeed. There's enough interest in an AI image generator that can create consistent characters that someone is bound to figure out how to do it eventually.
@@vainezaiven6677 Let me know when it's stupid simples... for now I'll stick with text stories :)
Since we already can create 3D characters from prompts it shouldn't take too long
@@missoats8731 Can you see my reply above? Thanks
We have it with midjourney character reference and scenarios pricing is ridiculous. What we need is for every character in a scene to be consistent. No one has that yet. Maybe SORA does.
The adventures of Santa Ciri are something we all need in our lives 😁. Thanks for the great summary vid.
Thank you very much for making this video! I've been looking for something like this for a long time and I'm finally getting great art for my original characters. I've been struggling with the side views, though. Did you use any negative prompts to make yours look so good?
Ive done some character turnarounds using auto 1111 but its very based on a single seed and prompt combo with openpose controlnet. Way more involved process and not as many strong outputs like this. Good vid.
That posing feature is insane I wonder how good you can get using those reference dolls and posing it to your content and then taking a picture of it
10:18 It would definitely be interesting to see how they can get the character posing system to be more flexible.
This is really cool. So I’m guessing we can use say Ai generated art from other applications, download them and then can upload them into Scenario? Or even use them as a reference to get the model in which the other applications do not generate what you are looking for. It will be really interesting how this will evolve, maybe be able to make scripts or animate what you want them to do.
Hi and many thanks for this video. I have a question. How can you achieve the 3 images on scenario (frontal, side back) when is not a human body, for example an animal cartoon? which reference image could I use for this case?
Schecter E/A!!!! I have that same guitar. Absolutely love it. You should make a video with it.
F-Hole Gang!! I love it too-- it's been a mainstay for years now! For sure a guitar that punches WAY above its weight! I feel like I can get practically any tone I need from it, and man is it smooth to play!
At some point soon I will! Next look at AI Music tools and I'll for sure get a little guitar in there!
Thx Tim! Keep it up for us Budz!
I have a lyric video project coming up soon and this could be EXACTLY what I'm looking for!! Good work!!!
When you did the prompt for ciri sitting in a tavern with a barbarian is it possible to use a second model, so you could use consistently two characters or even a group?
To be honest, I don’t remember. Sorry! But, I’m going to do a follow up video on Scenario, since they’ve added so many new features, so I’ll be sure to flag that!
You mentioned initially setting the reference image to have 0% influence, what does that mean? It still seemed to have a small bit of influence in the output since it did generate 3 character images in approximately the same poses as the reference image. Shouldn't setting it to 0% mean that it has no effect at all? Or does the AI still consider the pose but not any of the details or art style?
So, I think so? I noticed when I would turn up the influence that I would get a lot of muted colors in my outputs- which I think was referencing the grey tone look of the template model.
You know what might be interesting? Painting over the template model and using that as an image ref with the influence turned up.
Hmmmm, might have to try that today!
Yes! This is what I have been waiting for. Thanks man!
been messing with something like this in comfy using a workflow i found recently..as long as you lock the seed for your character/subject prompt and of course the same models etc, you get 99% the same character
the workflow uses two prompts: a background prompt for the scenery and a subject prompt for the characters… it then uses IPAdapter and condition combining to put the Subject into the background : background can change depending on prompt weights etc in the final image but the character stays very consistent.
even with keeping the seed randomized as long as the prompt is detailed enough it also keeps consistent.
Can you please share the name of the workflow? Thank you!
@@alincristiantiganus4137 Experimental Prompt Blending Workflow
is the original
@@alincristiantiganus4137 He said "in Comfy" !?
legend - as usual. thank you ⭐⭐⭐⭐⭐
Nice, going to share this in the PIKA channels , always getting asked for this :)
oh, thanks so much! I gotta stop by soon--Tell everyone I said Hi!
big announcement coming as well @@TheoreticallyMedia
Hey great tutorial, I had a question , can you load a png image character that you have drawn yourself and ai will make your character do different poses , I’m only new to this .
Im trying to make a graphic novel and need the ai to generate my character's consistently.
Would you recommend this way to do that?
Great video, Tim! Thanks for the heads-up on a new tool that can really make a difference. I'll contribute to your PDF and overall efforts, it's just a bit shameful that my local currency, Brazilian Reais, is so devaluated. 😢😮
Thank you so much! Oh, and please, never stress on the PDFs! You are all always just welcome to nab them. That is a bummer about Brazilian currency- particularly since you have Pao de Queijo there. In all honesty, those should just be world currency!
Hi, could you tell me before watchnig your whole video, if this works for photo-realistic consistent characters (seems you're using fantasy art style characters in the video). thanks!
Great video! One question, do you know if it is possible in Scenario, or any other generator, to create a character and name it. Then go on vacation for a week, and then come back and create more photos with that same character? If so, how? Thanks again!
Throwing appropriate shade on Nordic fairy tales. i love it.
There’s always a dark forest, and always something with teeth eating kids!
Consistent character is very easy in stable diffusion.. Just train the model. On top of that, many controlnets to put the model in any pose or outfit. Also face swaps can be done easily as well.
Training a model is a hassle.
Any good tutorials on how to do that?
@@Rayzorbladez indeed but I don’t like linking other channels on a person channel. It’s disrespectful. Just type it in the search box and many results will pop up. It was a big thing early 2023
Not everyone use High VGA
Hi, I followed the steps and my creations are line drawings and without color. Any ideas for a solution?
7:02 How do you 'scroll the image down'? I tried simply scrolling and holding down buttons like Shift and Ctrl while scrolling, but that doesn't work.
I have a mouse wheel and just rolled that. To be honest, I’ve never actually thought what that keyboard alt command would be.
@@TheoreticallyMedia Thanks so much. It doesn't seem to work like that anymore, as that's the first thing I tried, although I appreciate the quick answer!
Thank you for your Vedic. I’m trying now.😊
Is scenario not able to generate characters with braces? I can't seem to make it generate a character smiling with braces. I've tried prompts "braces", "teeth with braces", and "grinning with braces".
I have a problem creating the character turnaround.. It seems to have an inconsistency when comparing three different angles. How do I fix that? Could you give an example how to prompt the different views? Thanks.
Oh, I just had to do it a lot of times. You can try weighting the prompts with (((front view))), (((side view))) etc.
One trick I forgot to mention in the video is if you run across a correct angle character, but you don’t like the character, just download it and re-input that into the reference box and prompt from there. That seems to help.
@@TheoreticallyMedia thank you
🎯 Key Takeaways for quick navigation:
00:00 🌐 *Overview of AI Image Generation Challenges*
- Consistent characters in AI imagery have been challenging to achieve.
- Previous methods yielded janky results, making it a difficult area to cover.
- Introduction to a potential solution in the form of Scenario, a new AI image generator.
01:09 🎨 *Creating Custom Character in Scenario*
- Accessing Scenario's homepage and choosing a model from various aesthetic styles.
- Exploring sliders to control the influence of each model and using the prompt box for character details.
- Emphasizing the importance of front, side, and back view prompts for a comprehensive character generation.
03:28 🖼️ *Character Template and Initial Prompts*
- Utilizing provided character templates (male/female) or using a custom template.
- Initiating character creation by dragging the template, setting mode to image-to-image, and experimenting with influence.
- Providing specific but general prompts for age, hair color, clothing type, and focusing on front, side, and back views.
05:32 🚀 *Enhancing Image Quality with Upscaling*
- Exploring Scenario's upscale options (2x, 4x) and utilizing face restoration and enhanced upscale.
- Suggesting the use of an external tool (Magnific) for additional creative upscaling.
- Downloading the final character image for further processing.
07:12 🧩 *Training a Model in Scenario*
- Creating a new model in Scenario by dragging in images and adjusting for any horizontal images.
- Naming the model and providing additional information in the description box for better model understanding.
- Highlighting the importance of not altering advanced settings extensively for optimal results.
09:51 🎭 *Blending Models for Style Variation*
- Discussing the need to blend the generated model with another to avoid default poses.
- Demonstrating the process of composing with precision, adding multiple models, and adjusting sliders.
- Emphasizing the ability to test the blend before finalizing, showcasing various style variations.
11:44 🕵️ *Refining Poses with Control Net and Negative Prompts*
- Explaining the use of Control Net for refining poses and referencing real images for specific poses.
- Acknowledging that AI may still generate imperfect results and the need for iterative adjustments.
- Managing unwanted elements like elf ears through the effective use of negative prompts.
13:21 🎨 *Editing and Finalizing in Scenario*
- Briefly introducing the canvas function in Scenario for basic editing tasks.
- Mentioning the layers tab for experimenting with different generations.
- Expressing excitement for the potential development of Scenario as a platform for consistent AI character creation.
Made with HARPA AI
Didn’t someone already do this with Harpa? Don’t get me wrong, I do like the tool, it just is never accurate with the timestamps.
I was interested in images for a while but now video is so good especially with pika 1.0 website im waiting for consistant characters in video, i think thats when youtube is going to get much more exciting
Love the "This Is Spinal Tap" reference.
Can you train it to create consistent ‘backgrounds’ and consistent ‘objects’?
For example, I might need to pose my characters in a specific background that repeats throughout the story (and show that background from different angles).
Also, I might need to give my character a specific unique-looking sword, and have that sword appear consistent every time my character uses it.
It seems like AI image generation programs need to figure out how to do layers. That would solve a lot of the problems in terms of consistent backgrounds and other items.
You can-- I didn't really go into that in this video, so maybe I'll do another "Tips and Tricks"-- but yes: In your character training, you can also throw in things like locations and props. Just name them accordingly.
What I'm curious about is training two different models, with the same "look" and then blending them together. That might be a way of getting two or three locations and additional props. I'll have to try that out.
Thanks a million Tim!! I went to gunroad and I'm guessing the "I want this" button represents the download? Or am I missing something?
🎯 Key Takeaways for quick navigation:
00:00 *🎨 Generating custom characters with the Scenario AI image generator by selecting aesthetic styles and prompting for attributes.*
03:15 *🖼️ Customizing character templates involves choosing between male and female templates, adjusting sliders, and providing prompts for details.*
05:18 *🛠️ Upscaling and editing character images, exporting them for further editing, and cutting out different views for processing.*
06:31 *🧰 Training custom AI models by naming them, adjusting settings, and training with provided character views.*
08:18 *🔄 Blending models for unique styles, adjusting sliders, and testing for desired output.*
10:20 *📸 Using references and control net for specific character poses, experimenting with parameters, and understanding AI-generated poses' limitations.*
12:25 *🎨 Exploring advanced editing features, editing out unwanted elements, and leveraging layers for experimenting with different generations.*
Enjoy, not really sure if it will duplicate and reDraw already existing image, I hate the constant trial and error in that feature this is in March 2024 #
I wish we had you as a lecturer in the 90s, Your no bull, no clickbait and acknolegement , exctiment on this technology is infectious. Please keep up the good work. 😉
Haha. I’d just like to go back to the 90s for my baggy jeans and that one flannel I lost!
Thanks so much, and will do!!
@@TheoreticallyMedia Brother, I'm attempting to develop characters that look EXACTLY like that, thanks to you and this quick Scenario tutorial! More urban casual / hip-hop than grunge, per se, from the baggy clothes era of the 90's, but awesome just the same. lol
My only problem is the majority of the cool stuff I'm conceptualizing is happening right in Bing Image Creator using their free DALL E - 3 protocol in Co-Pilot's Chat GPT 4 and for as awesome as the results are (I'm even getting insane consistency with character generation through multiple prompt variations) I still can't recreate anything close to those characters in Scenario. And I want to REALLY badly! lol
Either way, thanks for the awesome no-nonsense tutorial just the same. Much obliged. You earned my subscription in the first 30 seconds! 😁
I don't seem to be able to scroll the image down for resizing while cropping to the 1:1
That seems to be a bit of a bug, I need to circle back to scenario and I’ll look into it!
Hello just trained 2 characters how can i use them to get 2 characters in a single image
Great video, but I already have my character designed I cant figure out how to use the image I have of it to generate other views?
Hi again.
Okay, so your name is Tim, right?
I'm enjoying playing around with Scenario, it does feel like it has some potential.
Various thoughts after several hours of experimenting:
I used one of your models, I used the female shape at three angles.
Still unable to get a generate model at completely side view though. what did you do to get completely side facing. most I've got is a women looking to the side and at an angle from the back.
The model I got from you, it's a greyish female humanoid with a sort of exoskeleton line path mapping out the contours. I went to google looking for other basic humanoid profiles and scenario could barely add features. But it had no problems making a real looking human from the model you provided.
I'm so confused by all this. I'm still basically in the first third of your tutorial.
So, I need to update this video (I'll do that soon)-- but an interesting and much easier way of doing this is to simply use Midjourney/Dall-E, or really any other image generator to create a character sheet for you. Just use the prompt (or something like): Character Turnaround, (style/character), Front View, Back View, Side View-- and then use those for your slices.
I don't know why I didn't think of that, but it does cut down on a lot of extra busy work.
I'm having a heck of a time getting Scenario to generate anything other than black and white illustrations of characters. I have the reference image set to IMAGE TO IMAGE and the reference to 0. And I've tried IMAGE TO POSE instead but still no luck. Every model I use, black and white illustration as a result. They are cool, but I wish I got something cool like in this video. Any ideas?
Got two quick ideas!
One (and this is kind of stupid, but it seems to work)-- if you have a photo editor, take the template and literally paint over the model with the colors you'd like the character to be wearing. you can go super basic here, we're just telling the model: "Hey, color here!" Also be sure to prompt for those colors and clothes "Black Leather Pants" "Blue Aloha Shirt" (That's my weekend look...)
Two: If you have a gen that you kind of like, with correct poses-- download that and re-input it in as an image prompt. Basically, you're buying a V.2 of the character template. That seems to help!
What if you want multiple characters in a panel ? What is the strategy then?
🎯 Key Takeaways for quick navigation:
00:00 🌌 *Introduction to AI Characters*
- Consistent characters in AI imagery have been challenging.
- Previous methods often yielded unsatisfactory results.
- Introducing Scenario, a promising AI image generator.
01:23 🎨 *Choosing Aesthetic Styles*
- Access the models tab to choose different aesthetic styles.
- Emphasis on blending and remixing chosen models.
- Explore sliders for controlling model influences.
03:00 🚀 *Creating a Custom Character*
- Select a template for a male or female character.
- Provide specific but general prompts for character details.
- Importance of prompts for front, side, and back views.
05:47 📈 *Enhancing and Upscaling Characters*
- Use the upscale button for higher resolution.
- Options for face restoration, photo optimization, and enhanced upscaling.
- Additional creative upscaling in external tools like Magnific.
07:26 🧠 *Training Your Own Model*
- Drag in looks for your character and adjust settings.
- Importance of providing detailed descriptions for character training.
- Caution against excessive tweaking in advanced settings.
08:58 🔄 *Blending Models for Better Results*
- Blend your trained model with other representative models.
- Adjust influence sliders and test the blend for satisfactory results.
- Demonstrates the importance of blending for diverse outputs.
10:32 🤸♀️ *Controlling Poses and References*
- Utilize image referencing and control net for specific poses.
- Addressing challenges in AI art and the need for multiple attempts.
- The significance of negative prompts to refine character features.
12:12 🎭 *Editing and Canvas Features*
- Introduction to the canvas function for editing images.
- A brief look at the layers tab for exploring different generations.
- Acknowledgment of Scenario as a dynamic and evolving platform.
Made with HARPA AI
Thanks. The timestamps are always a little off on these.
IPAdapter face ID + Insightface loader + IP adapter apply (face model you like + clip vision) does the same thing but better and works with all models, from anime to 3d to artistic to realistic
@0A01amir can you share the Comfy UI Workflow?
@@user-jc3sq9qc4y I give you the channel that has the workflow for 10 IPAdapters in one (the dude created IPadapter), th-cam.com/video/t2OBzV3UHv4/w-d-xo.html
delete those that give you bad results until you have 1. then add a UpscaleLatentBy with 0.5 denoise and boom, you will get an even better face and overall look of the picture.
Looks to be the same mix of LorAs and model blending available everywhere, but it does seem to get good results so maybe they've got a good handle on default config settings to use. I'd honestly go through this to record all the default values they give you and try it myself on a stable diffusion cloud install. Not a dig on them, providing a good UX and reliable results for non programmers has been something apparently really hard for most of these Stable Diffusion as a Service sites. The number of subscriptions I need for a complete workflow because each site might have one or two things they do better, has been growing to huge sizes.
There are also attempts at using AI to generate a 3d model from the turnaround sketches, I'm eager to see that. If you can get a simplified 3d model to manipulate into poses, you could then use that are your reference image for controlnet to feed back into your detailed whatever-style 2d image.
That’s a really solid take on the service platforms. Agreed on all point, particularly that each of them are offering a variety of services that don’t quite overlap, which- yeah, does end up bloating out a budget!
The 3D aspect is the one I’m keeping an eye on. I’ve spoken with a few of those companies already and the advice I always give is to make it akin to a video game photo mode.
Currently, they’re trying to balance features/UX with ease of use. Most of them could probably put together “Blender with GenAI” but without a simplified UI, most people wouldn’t bother with it.
I do think we’ll see the Midjourney of 3d this year though. It might even be Midjourney who nail it. We’ll see.
Another great video. I watched it through and then went step-by-step on a second viewing. Your character seemed to keep a far more consistent costume in all of the images you generated compared to mine, wonder if there was reason for that, or if you had some images where the costume wasn't much like your original?
Not going to lie, it takes some generating to get there. Prompting as well-- I would occasionally get her in a tan outfit, so I would reroll it with Ciri in her (((blue))) outfit holding...etc.
That seemed to help.
At the end of the day, this is still AI Art, y'know?
@@TheoreticallyMedia Good to know. I thought I was doing something wrong. The consistency does seem to be there more than with most methods, though, but on the free version I've got to the stage where I'm sometimes waiting for an hour and a half between rerolls... guess I've become spoilt with how accessible the tech is at the free level that waiting seems to take an age, but there's plenty to experiment with while waiting!
I couldn't really get it to work as shown no matter what I did. In particular color was an issue. The reference image would either make the whole thing black and white with maybe one item or two colored despite specifying color for each part individually or even just telling it to have color. I ended up greatly increasing guidance and got something much closer to what I wanted but could no longer get it to do the front, side, and back view. It'd be very helpful to see exactly what prompts and settings you used to help guide the changes I might need to make
So, try turning the influence down to 0 and see if that helps on the color front. 2 other quick ideas:
1) This is kind of stupid, but take any paint program, and literally paint the color of clothes you want on the template. That should tell the model to use those colors.
2) You can also try this if the 0 influence works, but not perfectly: Take the output and reload it in as an image reference instead of the template. That way the model is working with an "updated" version and has more details to work with.
Are the prompts in the PDF? I thought I put them in there-- if not, I'll try to add them in.
@@TheoreticallyMedia Oh, I didn't even think to check the pdf for the prompts. I'll check that. Thank you!
@@TheoreticallyMedia Unless you're talking about some of the images, I can confirm those prompts are not there. :(
Can you download the ckpt to use in SD?
How about posing multiple consistent characters together in a scene? Does this AI tool allow for multiple consistent characters to be posed together in a scene? When it comes to storytelling, it’s not enough to just have a single consistent character in the scene. The different characters will often have to interact with each other throughout the story, so it’s important to show these interactions between characters in the images that are produced.
I use 3D to pose the characters and stable diffusion’s IP Adapter and img2img to apply the character’s to the 3D img. Lastly, some polishing in Photoshop then I plug it back into img2img for a final pass. Not speedy workflow, but it works
@@tstone9151 Can you share the Comfy UI Pipeline?
Love your vids! Thanks for the update on AI news!! It's going so fast, I've already used AI for projects and I'll definitely use it more often from now on.
2024 is going to be SUCH an insane year for this tech! Buckle up, we're going for a RIDE!! (and thank you so much!!)
I don't know much about gaming. Are these AI characters for gaming images? I understand how important it is to create consistent images, like for a set of scrapbooking papers with a specific theme. Is this useful for things like that?
It is using obviously stable diffusion which you can use it unlimited and free on your own computer or on a dedicated server with a price of the server. But as a start point for understanding the basics good website to start with because it simplifies things. Thanks for the video
True! And I don't mind any of that! I think if you have the time/know-how/and the technical ability to get SD up and running, you totally should. But, if you want to just get started right away, Scenario (and platforms like it) are a great way to jump right in and start creating.
There is 100% a personality type who loves building and fine tuning the ultimate SD machine. But, it also takes a fairly heavy time commitment to learn to skills to do so, and-- well, let's face it, funds to do so.
Platforms are a good solution, since the cost is (relative to buying a machine) lower, as well as minimizing the time tuning new models.
That's the great part about current AI: Everyone Wins.
An artist looking at this can spot so many issues, but to the non-artist it looks great.
Agreed. I think the current best results come out when someone skilled with digital art programs can spot those issues and punch them up.
I don’t think that’ll ever change.
consistent characters is very needed at this point 🤘
It’s pretty much the holy grail of creative AI. And it looks like it’s basically here! 2024, let’s go!
Can I design my own character from scratch and use this to get more poses and place it in environments?
You can! You can also design your own environments to add to the training!
Nice! Custom AI concept art, here i come 👌🏾
I'm using the Male template and I've got a detailed prompt describing hair colour clothing colour and style all my characters are coming out white just like the Template I can't see no obvious solution any help would be much appreciated
At this point, I might say just to hope over to another image generator (any will do) and prompt for a character turnaround and bring that in.
Scenario has changed around a bit, I need to head over there for an update to this tutorial!
@@TheoreticallyMedia I'll give that a go thanks so much for Replying
Does it work with anthropomorphic characters only? Or I may create, for example, a consistent cucumber-based character?
So, I think you could? Instead of providing one of the humanoid templates, try doing a cucumber shape?
Technically, you should be able to train any object though.
Now I’m thinking about pickles and I want a really good one. Not that store bought trash. Like a crunchy one that’s been sitting in a barrel of brine!
Anyhow…yeah. It should work!
Love the humor Tim
Thank you! I'll print out your comment to shove in my kid's faces next time they eyeroll on a "dad joke"!
The characters generated are only shown in grayscale, like the reference image of yours that I used. I specified colors in the prompt, don't know what I'm doing wrong.
Did you turn the influence down? That seems to be a pretty key step. Also, interesting idea I’m going to try out later: taking the template models and painting over them (poorly) with a color scheme. I think that might nudge the model into some interesting directions.
I've been experimenting with consistent characters in mid journey by "casting" those characters with real life actors, and then providing prompts for hair, makeup and costume. It has worked surprisingly well.
wdym with casting?
@Unitedwestandseperatedwefall I tell Mid Journey to create an image of a particular actor by name, and then also tell it how I want them to look. The AI knows who the actors are (as long as they are famous) and renders a likeness of them.
@@richardbarcaricchio ahh yeah I get waht you mean when it's regarding "famous" people, there should be a way to imput pictures of (non famous) people (possibly in diff poses like in this video) and save them as a specific character in midjourney itself
Yup, I've used that method in the past as well. The one issue that is tough to get around is clothing, since outfits will often change. My old trick was to prompt for very basic attire, like "A Black Tuxedo" or "A Red Dress"-- that tended to keep things more or less consistent. But, moving beyond that doesn't really work.
I had done this a lot in the past to make comics, and it worked well from v3 onward. It seems to be broken in v6 from what I'm seeing at this stage. Prompting Tom Hanks will give you him maybe 10 - 20% of the time.
thank you for this been expirmaneting with wrighting now that we have editing tools the make it so my disabilites dont get in the way of wrighting
Great video. The morphing or lack of cosistency of characters was what bothered me the most with AI tools. This works very wll but when initially generating a model (bsesd on the one shared in this video) I often get one feet faccing a different direcion than he other as well as the face facing the wrong way or even two faces. This is in the the side view, front view and back viel look very well. I am rather new to this but does someone know some promts to avoind this?
Try weighting your prompts with (((front view))), (((side view))), etc. Also, negative prompt with things like “deformed anatomy”
I forgot to mention in the video, if you do run across a character that is “correct” but you don’t like the overall look of it, just download that image and use it as a new image reference. That seems to help in terms of poses and correct facing characters.
the reference image is making the character completely white. reference influence set to 0. how did u made it colored?
As a dm that cant draw im going absoloutely crazy with this
That’s awesome to hear! I’ve got some more Scenario stuff coming up soon!
Excellent work Tim. You have a great feel for what is good. Thank you for posting!
Thanks, this is cool! I'm waiting for Midjourney to add this kind of features, but it’s good to know that A.I. image generating is getting there: Consistent characters and the ability to pose them the way you want.
MJ has been saying this (maybe not exactly this method) is coming soon. But yeah. Totally agree! In a recent Office Hours, it was mentioned that they were working on something that was like ControlNet, but not ControlNet.
I think 24 is going to be a big year for MJ!
@@TheoreticallyMedia Great! I'm really excited about everything A.I. will bring in 2024! I dig your video's, keep posting!
5:30
How does Magnific compare with Topaz Labs GigaPixel in terms of quality of output??
Topaz for sure goes bigger, but Magnific’s real strength is that it can add an extra layer of prompt generation on your image as it upscales. Neat trick and can really make a good image pop.
That said, like any seasoning, it can also make it go wrong! Like Rosemary!
@TheoreticallyMedia
👋 what does that mean and how does it translate to increasing the fidelity of an image? Is it a layer of fixing something not originally generated with good coherence to the prompt?
I've not tried Magnific yet so I've done no comparisons but I'm curious to know from people who have compared the 2.
If an upscaler can increase fidelity or fix AI bloopers...... that would be a definite plus but is this what's actually happening with Magnific?
@@LouisGedo So, what it's doing is kind of like a creative enhancement? I think that's the best way to describe it. I think the best way to understand it is to see it in action...luckily, I made a video about it!
th-cam.com/video/VCP1R5-Zywc/w-d-xo.html
(You can ignore all the GTA stuff in the video, I just did it since the game was trending at the time and thought it would be a fun way to introduce a wider audience to our weird little AI world!)
@@TheoreticallyMedia
You've convinced me to do a personal comparison
Great video Tim. I’ve been struggling with this.