Get Accurate AI Headshots: How to Create Your Flux LoRA

Summary

AI Explored Podcast Summary: "Get Accurate AI Headshots: How to Create Your Flux LoRA"

Release Date: April 8, 2025
Host: Michael Stelzner
Guest: Jonathan Mast, AI Coach and Founder of White Beard Strategies

---

Introduction

In this episode of AI Explored, host Michael Stelzner welcomes Jonathan Mast to delve into the intricacies of creating accurate AI-generated headshots using Flux LoRA models. Aimed at marketers, creators, and business owners, the discussion offers practical insights into leveraging AI for personalized and high-quality imagery.

The Importance of AI Headshots for Professionals

Jonathan Mast emphasizes the growing significance of AI headshots in modern marketing and professional representation. He shares a personal example:

"Just literally today I got an opportunity to promote an event I'm speaking at in April... within about three minutes, I now had 12 photos of what looked like me at a college campus..." (02:20).

AI-generated headshots allow professionals to tailor their images to specific contexts, making them more relevant and appealing to their target audience. Whether it's for social media, sales pages, or event promotions, having versatile and contextually appropriate images enhances engagement and authenticity.

Building a Comprehensive Image Library

Creating effective AI headshots begins with assembling a robust dataset. Jonathan recommends:

"You need to have probably at least a dozen... a combination of images. Some can be selfies... torso up, waist up, full body shots..." (07:33).

Quality and diversity in the initial images are crucial. High-resolution photos with varied angles, clothing, and settings provide the necessary data for training the AI model to accurately represent the individual in different scenarios.

Photography Tips: Lighting, Angles, and Consistency

Michael Stelzner inquires about the technical aspects of capturing the right images:

"So the main thing it's going to key in on is going to be your face and your facial structure... I try to make sure you're looking at the camera..." (09:50).

Jonathan highlights the importance of:

• Lighting: Ensure your face is well-lit without harsh shadows. Natural, shaded areas are ideal for outdoor shots.
• Angles: A mix of frontal, profile, and slight turns can help the AI model understand facial structures better.
• Clothing Consistency: Wearing the same attire across multiple photos can help in maintaining a cohesive look in the generated images.

Utilizing Flux and FAL AI for LoRA Training

The core of creating AI headshots lies in training the Flux LoRA model using FAL AI’s platform. Jonathan outlines the process:

1. Account Setup: Create a GitHub account and deposit funds (e.g., $20) as FAL AI handles server hosting (16:18).
2. Access Flux LoRA: Navigate to FAL AI’s “Explore” function and select the Flux Trained LoRA model (17:04).
3. Upload Images: Submit the collected images either individually or as a zip file. High-resolution files are preferred for better training quality (20:05).
4. Define a Trigger Word: Choose a unique identifier (e.g., your name) to integrate into future prompts, ensuring the AI associates it with your likeness (18:54).

Jonathan notes the affordability and efficiency of this method:

"It's easy to do. Maybe I want to be in a different location... for probably less than $5, you can have a really great library of five or ten LinkedIn headshots that you can use." (06:38).

Enhancing Prompts with Custom GPT: Felix

To refine the image generation process, Jonathan introduces Felix, a custom GPT tool accessible via [e-mail]. This tool assists in crafting effective prompts tailored for the Flux model:

"Felix then creates five unique prompts... It knows how to say things in a manner that Flux likes." (26:43).

Using Felix, users can input a general concept, and the tool generates detailed prompts that incorporate the trigger word, enhancing the accuracy and relevance of the generated images.

Generating and Refining AI Headshots

After setting up the model, the generation phase involves:

1. Running Inference: Input the enhanced prompt with the trigger word to generate images. Jonathan mentions:

   "Every image is going to cost you three to five cents." (33:32).

2. Adjusting Settings: Users can modify aspect ratios (square, portrait, landscape), guidance scales (typically between 10-15 for balance), and output formats (JPEG or PNG) based on their needs (30:12; 32:00).

3. Review and Iterate: Not all generated images will meet expectations. Jonathan advises:

   "Expect that's part of creating images with AI. I toss at least 50% of the images that I've created because it's messed something up sometimes." (33:41).

Through iterative refining—regenerating images with adjusted prompts or settings—users can build a library of high-quality, personalized headshots.

Refining Models with AI-Generated Images

For those seeking even higher accuracy, Jonathan discusses the possibility of retraining the Flux LoRA model using selected AI-generated images:

"We took about 24 AI generated images and I found it made a bit of a difference... it was worth the time and effort." (37:06).

This secondary training can fine-tune the model, ensuring that favored features are consistently represented in future generations.

Ethical Considerations and Disclosure

Throughout the episode, Jonathan addresses the ethical implications of using AI-generated images:

"If you're representing something that's not accurate, then you want to make sure that you're disclosing that." (38:32).

He advocates for transparency, especially when images depict scenarios or achievements that are fictitious. Proper disclosure maintains integrity and trust with the audience.

Creative Applications and Future Potential

Beyond standard headshots, AI-generated images can be creatively utilized for various purposes:

• Animated Videos: Using Flux LoRA, users can animate static images to create dynamic visuals for marketing or content creation.

  "It's a great way to have a little bit of fun and create some B-roll video or background video..." (40:34).

• Style Customization: Training the model on specific artistic styles (e.g., Pixar) allows for diverse and engaging visual content.

  "I can take an image of you and render that in the style of a Pixar image." (39:40).

Conclusion and Final Thoughts

Michael urges listeners to experiment with AI-generated headshots and share their creations on social media, tagging both him and Jonathan. He also encourages joining Jonathan’s Facebook group, "AI Prompts for Entrepreneurs," which boasts nearly 360,000 members.

Contact Information:

• Jonathan Mast: jonathanmast.com Linktree
• Facebook Group: AI Prompts for Entrepreneurs

Michael concludes by highlighting the potential of AI to enhance professional success and invites listeners to explore further episodes of the AI Explored podcast.

---

This summary encapsulates the key discussions and actionable insights from the "Get Accurate AI Headshots: How to Create Your Flux LoRA" episode of the AI Explored podcast. For detailed notes and additional resources, visit Social Media Examiner’s podcast page.

Transcript

[Michael Stelzner]

Hey, before we start today's show, if you want to accelerate your AI learning, I have a solution for you. Become a member of our AI Business Society. You'll join me as we go deep with live AI training each and every month. Imagine crafting more persuasive content, creating stunning images and automating those time consuming tasks. It's all possible when you join the AI Business Society. Go to socialmediaexaminer.com AI and join today. Welcome to the AI Explored podcast, helping you put AI to work. And now, here's your host, Michael Stelzner. Hello, hello, hello. Thank you so much for joining me for the AI Explored podcast brought to you by Social Media Examiner. I'm your host, Michael Stelzner, and this is the podcast for marketers, creators and business owners who want to know how to put AI to work. I'm really excited about today's show. I'm going to be joined by Jonathan Mask and we're going to explore how to create accurate AI headshots that look just like you. If you have seen people creating these headshots that look ridiculously high quality and you've wondered how to do them, this is the episode for you. By the way, if you're new to this show, be sure to follow us so you don't miss any of our future content. Let's now transition over to this week's interview with Jonathan Mast, helping you simplify your AI journey. Here is this week's expert guide. Today, I'm very excited to be joined by Jonathan Mast. If you don't know who Jonathan is, he's an AI coach and founder of White Beard Strategies, a consultancy that uses AI to help businesses save time, deliver more value and become more profitable. His course is AI Prompting Mastery, and his newsletter is the AI Prompting Insider. His very successful Facebook group is AI Prompts for Entrepreneurs. Jonathan, welcome back to the show. How you doing today?

[Jonathan Mast]

I'm excited to be here. I'm doing great. Thanks for having me back.

[Michael Stelzner]

I'm super stoked that you're here. Today, Jonathan and I are going to explore how to create AI headshots that look exactly like you. And I'm sure people that are listening have seen these. They're really, really cool. We're going to get into it. By the time we're done today, you're going to be able to figure this out all on your own. Now, before we get into the details, Jonathan, why should marketers, entrepreneurs, creators care about creating AI headshots of themselves? Let's kind of focus in on kind of, like, the uses.

[Jonathan Mast]

Well, I'll give you an example I use today. So just literally today I got an opportunity to promote an event I'm speaking at in April. I'm going to be at a college campus. I don't have any pictures of me at a college campus. And within about three minutes, I now had 12 photos of what looked like me at a college campus dressed in a navy blazer and a blue oxford and a baseball cap with the Marvel Avengers stuff on it. And again, I did all of that in just a few minutes. And now when I post that it's applicable, it's relevant. It's not me with my typical headshot. It's not me standing in a studio. I'm literally standing on a sidewalk. It will look like a university. And that's going to appeal and be relevant to the marketing message.

[Michael Stelzner]

So, okay, that's really cool, but where else could we use these kind of images? Because I know you do these all the time and you take it for granted, but not everybody's gone down this rabbit hole, you know, Know, Fair enough.

[Jonathan Mast]

I use them on social media to make sure that I've got relevant things for social media. You know, if I need to have a picture of me looking a particular way the other day, my. My staff, I did a video and they said, we need a picture without your hat and with your hands on your head and you looking excited. Well, that's hard to come up with really quickly. Now we can do that. Not a problem. Easy to do. Maybe I want to be in a different location. Maybe I want a different type of clothing on whatever the case may be. Literally the other day, somebody was playing around and they said, for fun, can you do this and put you on a tyrannosaurus riding on its back? And yes, I could. Some of that makes sense to you. Some of it's just for fun. But a lot of it, again, is really about making sure when you're sharing images online, especially for marketers and business people, that those images are relevant to the discussion that you're having.

[Michael Stelzner]

I like that. When we were prepping for this, I want you to share the. The vision board story of Lindsay, who was a consultant. Do you remember that story? Can you share that a little bit?

[Jonathan Mast]

I do, yeah. So Lindsay's one of my coaching students, and she did something absolutely amazing. So she's big in the vision board. She teaches people how to do vision boards. And if you've ever done that before, you know you're going to run out and you're going to go ahead and find a magazine or find an image online and you're going to grab that. But what if you could create a vision board that included pictures of you doing what was part of your goal? So imagine I want to take a trip to Paris and I create a picture of me standing at the Eiffel Tower. Or if I want to go downhill skiing in Vail and I get a picture of me doing that, that I can add to my vision board that is actually going to resonate with me. So. And help me achieve those goals.

[Michael Stelzner]

I love that. So the obvious applications here are you can put these on sales pages. You can use them as your profile photos, as you mentioned, you can use them if you're speaking at an event and you want this to look relevant to the audience. So the applications are kind of endless and the benefits, it seems to me, and I'm just, I want you to like, add on to what I'm about to say. Not everybody has what they would call really great pictures of themselves, right?

[Jonathan Mast]

Absolutely.

[Michael Stelzner]

And not everybody can afford to go out and hire a professional photographer. Does this help solve for some of that?

[Jonathan Mast]

It absolutely does. You know, I think we've all been in that spot. We don't have the right picture or the picture is as of a couple of years ago, and maybe it's not relevant. So, yeah, with this now, you can literally take pictures from your cell phone selfies that you can take, upload them into the tools that we're going to talk about, Michael, and you can then train that model to. To create images that represent you that look like you in almost any scenario that may be just a professional headshot or I place myself in a photography studio on a clean white draped background or anything else for that matter. And yeah, it gives us that ability to do so for a few dollars in costs as opposed to hundreds or thousands of dollars.

[Michael Stelzner]

Now, I've seen these kind of images in the past and they tend to look a little artificial and cartoony. Are we beyond that now? Do they look real now?

[Jonathan Mast]

Absolutely not. All of them do. And it's probably an important thing to reminder when we create images like today. When I created them, I threw away about half the images because it wasn't quite right. But I was still able to get the images that I needed and do so in a few minutes and for a total cost this morning of less than a dollar. And again, to have that ability to do that in my mind is, is just amazing. It may just be you need a new LinkedIn headshot. Your LinkedIn headshots from three years ago. You can learn how to do this. And for probably less than $5, you can have a really great library of five or ten LinkedIn headshots that you can use.

[Michael Stelzner]

Okay, so let's start from the beginning. What are the first things that we need to be doing? We need to be thinking about in order to be able to create these AI headshots?

[Jonathan Mast]

Well, before we do anything, we need to understand that you're going to need probably at least a dozen. I generally recommend about two dozen images of yourself, and they should be a combination of images. Some can be selfies that you've taken with your phone. Some might be hand your phone to your spouse or your family member or a friend and go, hey, just literally take some pictures of me. Torso up, waist up, full body shots. You want to mix them up a little bit. And then you may have maybe it is a professional shot from two years ago that you want to put in there. And maybe you were at an event and you had a picture and you can get the other person out of it. So it's just you. A combination, though, of about. In a perfect world, about 18 to 24 images that you feel represent you. Different clothes, different stances, different perspectives, all as long as your face is showing that you are ready to upload, because you're going to need that training data to train the model on who you are.

[Michael Stelzner]

Okay, so I got a couple of questions on this. There are some people that, like yourself, have. You have a baseball cap that you wear, right? Right now it's Avengers. I don't remember it being Avengers last time. It might have been something totally different. So talk to me a little bit about how important it is that we get the clothing right. And also any tips on, like, the angles that we ought to take the shots at.

[Jonathan Mast]

Great question.

[Michael Stelzner]

Let's get into that a little bit.

[Jonathan Mast]

So I wear glasses. So I made sure that the images that I had had my glasses on because I wanted my glasses to portray realistically. Obviously, as you can see, if you're listening, you can, but I've got a long white beard. I wanted to make sure that was portrayed effectively. I didn't. I'll remove the hat here, but I didn't put any pictures of me with a hat on. I took pictures of me without the hat because it's easy to add a baseball hat in. Ah. I took pictures with me in different shirts. And again, some that were, you know, face shots, head and shoulders, like you're seeing online. Others that were Torso up and then a couple of full body shots as well, so that I can understand that I'm not built like a Greek Adonis. And if it's going to create a photo of me, it should probably be a little bit more physique accurate.

[Michael Stelzner]

I like that. So what about the angles and stuff? And also to our female audience, is there anything that they need to be thinking about when it comes to the shots that they're taking? Should they? Because you know how a lot of women change up their hairstyles and stuff, right?

[Jonathan Mast]

Yeah. And that is going to matter. So the main thing it's going to key in on is going to be your face and your facial structure. But obviously, if you do a lot of different hair designs, that could make it more difficult. You'd want to upload again, more samples of those if you want to do that. And then you're just going to need to remember when you're prompting it, which we'll get to in a minute, that you describe that appropriately. Facial angles. I try to make sure you're looking at the camera, assuming that's the type of images that you're going to want to be creating. If you only wanted to create side shots, then for those that are listening, you couldn't see me turn sideways in my chair. But then you would create and upload images of you in that profile perspective instead. I made sure that I was standing straight on the camera. I turned a little bit and leaned in on a couples. I made different things, but I tried to make sure my face was. Was visible in all of those images because I knew that's what it would be modeling to try to recreate.

[Michael Stelzner]

Yeah, it's intriguing because anybody who has ever had a professional photo shoot done, maybe you think back to your wedding, they have you try all sorts of different angles. Right. Like they have you tilt your body to the side, but have your face. Face the camera sometimes dead on shots. Do you recommend, like, as far as the angle of the picture? I don't know, like you could take it from dead on versus up a little bit, versus down a little bit. Does that kind of stuff matter to the AI at all?

[Jonathan Mast]

It's a great question. I've not tried enough different angles. I readily admit I originally trained the model that I use on about 24 different images. And I've been so pleased with the output that I've never went back to try different things. At that point in time. It just worked super well. And again, most of the images that I did, I had a photo shoot that I'd done last summer that I uploaded a couple images from. Most of mine were selfie images or images that I had from events that I was participating in where, you know, Michael, you and I might be standing next to each other, and I would just use Photoshop or something to pull you out of the image and just put my image in.

[Michael Stelzner]

Now, what about the backgrounds on the images? How important is this, like, inside, outside kind of stuff?

[Jonathan Mast]

Lighting is probably more important than the background because the background's really going to be defined by the prompting that you use. So the key is that your face is well lit enough that it can see it. So, you know, I've got lights on right here. If I turned off my lights, that's going to make it harder to see the details in my face, the details in my beard, the details on. On my glasses. So, you know, a great thing is good light, but you also don't want to necessarily be outside in the bright sunshine. As I look outside right now, if I was there, I'd be squinting and it would be overblown. So balancing that out, if you're outside, maybe find a shaded area that's still got good natural light. If you're inside, try not to stick with a dark room. Get something again where your face is reasonably well lit.

[Michael Stelzner]

Okay, so the goal here is to end up with about two dozen shots.

[Jonathan Mast]

Correct.

[Michael Stelzner]

And it sounds like we don't need to love them all, we just need to like them all. Is that kind of what I'm hearing you say?

[Jonathan Mast]

As long as they represent you in a. In a fashion you're comfortable with. Again, it's training on your face. So keep in mind that's the part that we're really trying to train it. Hairstyles, to a certain extent, beards, if, for men, if you have them, to a certain extent. But it's really training on this structure that is as on mountain lighting. And for those you're listening, my head, my face and my beard, that's a part of your body that it's really going to be looking at. The other body shots are helpful, but the main part is your face.

[Michael Stelzner]

Okay, let's say we end up with 25, 26, 30, 24 shots. What do we do with all this?

[Jonathan Mast]

So the next thing, we're going to be using a software called Flux Flux. And that is not a software you can just go to a website and subscribe to by normal terms because it's actually software that's available to download. So we're going to recommend they actually Go out to the developer's website and I use a site that they recommend that's called FAL AI and there are a lot of models that I can do on that. But Flux is one of those models that works really well and it's why I recommend that particular site.

[Michael Stelzner]

Real quick, you mentioned the developer without saying the name of the developer. We both know there's a lot of people out there masquerading as the developer. But who is the actual developer if they were to go look that up and see the recommended tools?

[Jonathan Mast]

So it is actually Black Forest Labs.

[Michael Stelzner]

Okay.

[Jonathan Mast]

Which makes it even harder because if you Google Flux, you probably will not find a link for Black Forest Labs. You'll find a link for a lot of people saying they offer Flux. They do, but it's not the model you probably want.

[Michael Stelzner]

Okay. My understanding of Flux is it's kind of an open source model or it requires API access. It's not like a piece of software that we're used to using like ChatGPT or Claude. And as a result of this, there's a lot of services that offer this. F A L A I is the one that you are recommending and you use their Flux model. Can you just talk to us a second about why Flux? What is it about Flux that makes it so powerful?

[Jonathan Mast]

You can do this with a few other models as well, but Flux is the easiest to establish and train. What's called a lora, an L O R A or a low resonance model of your face. There's tons of graphics out there, but. Or graphic tools out there, but very few of them allow you to actually upload data. And then it creates a training file that resembles Jonathan or that resembles Michael. And that's then what I use to create those additional images.

[Michael Stelzner]

Well, and I'll also add that I use Flux Pro Ultra, I think it's called, to generate realistic images of people. And I feel like Flux is one of the best models for creating photorealism. I don't know if you agree with that or not.

[Jonathan Mast]

I do, 100%.

[Michael Stelzner]

Yeah. That's the other side of it is like the model is trained, unlike Dolly, which always feels to be a little bit less realistic, more artsy. Right. Wouldn't you agree?

[Jonathan Mast]

Yes. Yeah, 100%. Yeah.

[Michael Stelzner]

So, okay, so we go to this FAL AI website and what do we need to know when we get in there?

[Jonathan Mast]

What you need to know when you go in there is that in order to use Flux, because it is like you mentioned, it's a model that can be downloaded and the reason I like FAL is that they, they handle all the server hosting and everything else. You don't need to worry about any of that. But it does cost a few pennies in order to use it. And so when you go there, you're going to need to set up an account. They give you a couple different ways to do that. I recommend using a service called GitHub.g I t H U B. You will have to deposit some money into that account. GitHub is a multinational company. Whether you've heard of it or not, it's very reliable. And I deposited 20 bucks when I got started. I said, okay, let's put $20 in.

[Michael Stelzner]

Aren't they owned by Microsoft? I feel like they might be.

[Jonathan Mast]

They might be. I don't honestly know, but they might be.

[Michael Stelzner]

Yeah. Just a real quick tangent. GitHub is really like a big repository for code, for lack of better words. I mean, like a lot of developers, technical people understand GitHub, non technical people have probably heard GitHub and their eyes have rolled backwards. Is setting up an account relatively simple?

[Jonathan Mast]

As easy as setting it up on Facebook or anywhere else? You literally just give it your email address, then they're going to ask if you want to deposit any funds. You say yes, because you are going to have to pay as you go for this model. And again, I recommend 10 or $20, and that'll last you a long time.

[Michael Stelzner]

Okay, so we've set up a GitHub account, we've deposited $20. What do we need to do as far as this FAL AI side of things?

[Jonathan Mast]

So we're actually going to start at FAL AI and then it's going to give you that option to fill it out. And so as soon as you do your account at GitHub, it's going to come right back to Fal AI. As long as you start at Fal AI, you're good.

[Michael Stelzner]

Real quick, should we have had the GitHub account already established before we start to. No. Okay. It's not necessary.

[Jonathan Mast]

Okay, not necessary. If you've got one, you can use it, but no, you don't need to. And that's why if you start at FAL AI, it's really easy.

[Michael Stelzner]

Okay.

[Jonathan Mast]

And then once you're signed up for your account, deposited your 10 or $20, it's going to bring you right back to FAL AI. And then we need to use their Explore function. So just a quick disclaimer, FAL has lots of different software products, visual video image products they make available, and you're simply going to search for what's called Flux. Flux Trained Laura. When you do that, it will bring you directly to a link that will allow you to train your own model.

[Michael Stelzner]

Okay, cool. So we've located the Flux Trained Lora feature, for lack of better words, inside of FAL AI. What do we do next?

[Jonathan Mast]

Once we do that, you're going to be presented with an input form because there's not going to be anything trained for you on your account at that point. And you're going to have the opportunity to upload those images that we just talked about. You can upload them one at a time or as a zip file.

[Michael Stelzner]

Okay. When we were prepping for this, we talked about Fast Lora training. Are we getting ahead of ourselves here or where are we with that?

[Jonathan Mast]

Nope, that's right where we're at.

[Michael Stelzner]

Okay.

[Jonathan Mast]

In F, A, L, it's called Flux Lora. It's L O R A for those that are spelling dash Fast Trini, and that'll get you right to where you need to go.

[Michael Stelzner]

And by the way, we set up a short [e-mail] L O R A will automatically get you there so people don't have to worry about, like, remembering all this.

[Jonathan Mast]

That's perfect. Yeah, that's perfect. Yeah. So then you literally are presented with an opportunity to upload your images and then to choose what we call a trigger word. And this is a phrase or a word that you're going to use anytime you use this model in order to replace the picture with yourself.

[Michael Stelzner]

What do we need to know about that? Give us an example what you mean by a trigger word.

[Jonathan Mast]

Mine, I just put my name in. So I use my trigger word. Is my name Jonathan Mast? And it's all one word together, no spaces. And so now, in the future, when it goes to train, because the next step, we're going to literally say to start the Training takes about 15 minutes. It's going to analyze all your images. It's going to create all the map. By the way, that training currently, as we're recording this costs $2. That may change over time, but currently that's a $2 expense every time you do it. And then it presents you on the right side of the screen with a hey, here's an alphanumeric code that represents your model, and it's super easy. You then just click a button that's called run. And when you click that, you can enter your prompt, put in your trigger word that you gave it into your prompt, and you're Starting to create images that resemble you.

[Michael Stelzner]

Okay, so let's talk through a couple of these steps. First of all, we've gone to, after we've set up the account to social mediaexaminer.com Laura, just to get there quickly, we've uploaded our images and as you mentioned, they could be a zip file or you could just manually, I guess, upload them. And then as far as the size of the images, is there any restrictions on the resolution or size of the images in your professional opinion?

[Jonathan Mast]

I'm sure there are some. Not that I'm aware of. You want to make sure that you're uploading those high quality images as you have versus low res images. So don't take an Image necessarily. Your LinkedIn profile that you have, that's from LinkedIn, if you save it, it's going to be a very optimized file that will not work as well as the original file that you have from your camera. Upload the highest resolution file that you have.

[Michael Stelzner]

Well, and for anybody who has iPhone, you know, they have their own file format called Heic or something crazy like that. Does it work with those kind of images or do we have to convert these over to JPEGs or something?

[Jonathan Mast]

I am pretty sure it does work with HDIC. As long as you put them in. I'm 99% sure it does.

[Michael Stelzner]

Okay, so this trigger word, does it have to be something completely original that no one else is using?

[Jonathan Mast]

No, it doesn't because it's going to be combined with your Lora training model. So it's basically, if you can imagine, it's grabbing the model and then saying, what's the trigger word for the model? The model is going to be trained on that. So I could use Michael as my trigger word and that would be fine. It doesn't need to be unique.

[Michael Stelzner]

Got it. It just needs to be something that you're going to be able to easily.

[Jonathan Mast]

Remember that you can remember. Yeah.

[Michael Stelzner]

Okay, so looking at the interface here, there are some other options that we have. One of them is called In Style. Talk to me about whether or not that is important. And then also we talked about masks and whether we need to create masks and also steps.

[Jonathan Mast]

I leave everything by default when I go ahead and run my first training. So when I'm training that Lora, I leave everything by default In Style is turned off. Create Mask is turned on. And steps as we record this are at 1000. So I just would recommend leaving those at the defaults. That'll work just fine. The InStyle is probably the only button that you may want to use at some point. And that means instead of training it on your face or your image, you're training it on the style of image. As an example of what that means, I took A number of 3D Pixar Disney animated style images and I uploaded those totally different characters, totally different pictures, but they were all done in the style of what you might expect in a Pixar movie. And I then clicked the style button and my trigger word then was 3D cartoon. So I can use that style. It will then create any image in that style of image. So the first thing you want to do though is likely train your face. So you don't want to have that on in your first training in most cases.

[Michael Stelzner]

Now, when we were prepping for this, Jonathan, you recommended possibly increasing the steps from a thousand to fifteen hundred to two thousand. Has any of that changed since we last were prepping for this and what are the benefits? To take it up to more steps, potentially.

[Jonathan Mast]

So the more steps just is essentially the more reasoning it's going to do on your files, the more analysis of the file it's going to do. In my experience, a number between 1,000 and 2,000 is fine. They have changed their default not too long ago to a default of 1000. That should be ample. I wouldn't bump it up. You can bump it way up if you want to, but to like 10,000. Two things are going to happen if you do that one. It's going to take longer because it's got to process more and I'm not sure. But it may actually charge you more money too. I'm not sure about that.

[Michael Stelzner]

I would imagine it probably would. Okay, so so far we've, we've uploaded all the images, we've clicked the InStyle if we want to mod model the style and we've put our trigger word. How long does it actually take? What happens after we hit the whatever.

[Jonathan Mast]

Button, you hit the start button at that point and it, it starts analyzing. And that, in my experience, takes 10 to 15 minutes for it to do that and to build out. You'll know because as you're looking at the screen when you do this, when the new model is done, there's a training history box. It's on the right side of the screen and as soon as it's done, your trained Laura will pop up in the training history box.

[Michael Stelzner]

Okay, cool. So once we've trained this thing, what's the next part of the process?

[Jonathan Mast]

So once you've trained it, you would go back to that same page the next time you wanted to use it, which could be immediately or it could be a day later, Not a problem. And under the training history, again, you're going to see the Lora that you created, and there's a button that says run inference. And you just click on that button, and as soon as you do that, then it opens up a new schedule for you. And on the left side of the screen, or it's stacked if you're on a narrow screen, but you'll see the spot where you put your prompt in. So, like any AI image tool, we then just prompt it with the type of image that we're looking for, and then we make sure that that trigger word. So in my case, my name, Jonathan Mast, that trigger word needs to be part of the prompt at the appropriate spot.

[Michael Stelzner]

Okay. That's the reason why the trigger word is there, because it's going to be in the prompt. So you need to pick something that's unique that you would not normally use in a prompt is what I'm hearing you say, right?

[Jonathan Mast]

Yeah. So, for example, the images that I created of me on the university campus, I had a prompt that I had chatgpt help me write that was basically a stocky man with a long white beard and glasses standing on a college campus in Arkansas, because that's where I'm speaking at. I replaced that when I pasted that in. Instead of that, it was Jonathan J. Mast. That's my term as a stocking man with a long white beard and glasses standing at a university campus in Arkansas. By putting that, that trigger word in there, that's what allowed it to pull my image and recreate me in that image.

[Michael Stelzner]

Does it matter where you put the trigger word in the prompt?

[Jonathan Mast]

You want to make sure it's still logical so that it makes sense for AI. So as you're reading it, instead of, I could have just put Jonathan J. Mask with a long bite beard or something like that. But I often use one AI model, ChatGPT, to help me create prompts for the image models. And in that process of doing so, in fact, I think we have one we're going to share with everybody later, Michael, on that. In the process of doing that, that makes it easy, really, really simple for me to just find the spot where it references, you know, a man or whatever. And I put my name in there. That trigger word goes in there?

[Michael Stelzner]

Yeah. So we set up a short code called socialmediaexaminer.com Felix F E L I X and that goes to your custom GPT why don't you explain a little bit about what your Felix custom GPT does?

[Jonathan Mast]

So when we're trying to create images, Michael, saying it in the right way makes a big difference and there is a bit of trial and error involved. I want to encourage people, if you haven't done AI image imagery before, it's not a perfect science. So not every prompt is going to turn out exactly the way we want. Felix is a GPT. It runs on any. You just need a free chat GPT account and you can put in the concept of what you're looking for so you don't have to put in all the details, you can just put in the concept. Unlike this morning, a man standing in a university campus in Arkansas. That's what I put into Felix. Felix then creates five unique prompts. It's been trained on Flux, so it knows how to say things in a manner that Flux likes. And then you take those prompts, copy and paste them into the input box in fal, AI add your trigger word, and then there are some options you can select, like how many images in that, and then you are off to the races.

[Michael Stelzner]

Okay. Speaking of prompts, outside of using your custom [e-mail] Felix and the only reason we're saying that is because it's a really long URL, this is going to get you there faster because OpenAI doesn't do a great job with short URLs. What other kind of tips do you have about prompting? Because you had mentioned that you can add hats and stuff later. Like, how do we do stuff like that?

[Jonathan Mast]

So, you know, one of the things I did is I just, when I was in Felix, I said I wanted to be a man with a long white beard wearing glasses and a hat, a baseball cap that had Marvel Avengers information on it because that happened to be the hat I had on today. You may go. Well, Jonathan, we trained it on your face. Why did you need to tell it glasses and a long white beard? I did that because I've just learned that sometimes. And you'll do this as you learn sometimes different facial features that we have that are unique. It helps to remind the model to include those. So if I don't say with glasses, about 80% of my images come through with glasses and about 20% of them I look different because my glasses aren't on. And so I've learned just to put glasses. And I've learned in my case, because I have a beard and it's a longer beard, if I don't specify long beard. Sometimes it gives me a beard that's cut right to the chin that doesn't look as much like I do. So those are things you'll learn over time that the model may need to be reminded of just to give you the best possible image.

[Michael Stelzner]

Will the Laura know what your glasses look like?

[Jonathan Mast]

It will, in fact. Yeah. So if you train the model with my glasses, the vast majority of the images that I do of myself, the glasses look I identical to the glasses that I wore. Now, all of my training images had the same pair of glasses. If I'd had different glasses, that might not be the case. But I had the same exact glasses on in every image. And when I create them, they look just like my glasses.

[Michael Stelzner]

Okay, so so far what we've done is. I'm just going to start from the top. We've created a GitHub account. We've gone to FAL AI. We've funded it with like five or ten dollars. And then we've gone ahead and we have trained up our Lora using the FAST Lora training model. We've uploaded the images and now we're at the point where we're about to generate images and we've learned about enhancing the prompts, specifically using your custom GPT call Felix. Once we get that enhanced prompt, we're going to use our trigger word. Right? Like Michael Stelzner, Jonathan J. Mast. What else do we need to know when we're rendering images from the Flux Lora inside of FAL AI?

[Jonathan Mast]

So there are a number of options. If you're not sure, you can literally just click the run button that's underneath where your prompt is, and it will create the images using the default. There are some things that you can play around with, though, that can have an impact on it. For example, you can choose a different aspect ratio of your image. Do I want it to be portrait or landscape style? Square. You can make that choice. There are a number of steps that you can have it use. I've. I leave that at default just for the reminder. It's just leave the steps. They're called inference steps. Leave that at the default.

[Michael Stelzner]

Real quick on the aspect ratio, does it matter what the aspect ratio of the original. No shots were that we took?

[Jonathan Mast]

It does not.

[Michael Stelzner]

So we could take them vertical, horizontal. It really doesn't matter, is what you're saying. Okay.

[Jonathan Mast]

Correct.

[Michael Stelzner]

Okay, good.

[Jonathan Mast]

Yeah, good question.

[Michael Stelzner]

So aspect ratio formats are what, square, vertical and horizontal? I mean, for lack of better words, essentially.

[Jonathan Mast]

Yep, yep. You've basically got square portrait and then Landscape. And then you can also put in a custom one if you want. So if you had a very unique one, let's say for a Facebook banner, LinkedIn banner, you could put that in as well and it will then match whatever, whatever you put in.

[Michael Stelzner]

Okay, cool. What are our other options?

[Jonathan Mast]

There is a seed which very rarely are you going to use. I've actually never used the seed. The seed would be if I was going to give it another image that I wanted it to reference. So in general, I'm not doing that. So the next one is guidance scale. And this is one you're going to want to experiment with. So the guidance scale starts off. They call it cfg. It starts off relatively low. It's a numeric number from 0 to 25, I'm sorry, to 35. And I generally recommend going about a third to the halfway up the scale. That means somewhere between about 10 and 15 for the guidance scale. Now, this number may change over time, but I've found as it's changed, staying about a third to a half of the way up the scale is a really good spot to be.

[Michael Stelzner]

What the heck does this thing do, this guidance scale thing?

[Jonathan Mast]

So the lower the number is, the more creative freedom you're giving the model to take with the, the training data that you gave it. So the pictures will likely look less exactly like you. So I may end up with a really crazy beard. I may end up with a skinny body instead of a fat body. You just, you never know what could happen. As I get higher, you may go, well, Jonathan, why don't we want to max it out so it always looks most like us? The, the more, the higher we get. Then it starts changing things like our skin and things like that so that they. The best description I can give you is they look more plasticky. They don't look as realistic. I find once we get much over halfway point, then aspects of our face in particular will start looking less realistic.

[Michael Stelzner]

Okay, so what I'm hearing you say is right now somewhere between a third to the halfway mark. Experiment with that and see if you like the output. And if it's getting way too creative, then bump the scale. Is it just like a number you have to put in there?

[Jonathan Mast]

Yep. There's a slider that you can move and just adjust as you move the slider or you can enter an exact number in.

[Michael Stelzner]

Okay, what else do we need to know there?

[Jonathan Mast]

The last option is the number of images. I should say the second to last, it's number of images. You can go one to four Images. If you do four images, you're going to pay for four images and you pay for whatever you create here. So if you're on a budget, maybe one at a time is better. If you're in a hurry, like I tend to be in a hurry, I then do four at a time because I. I just don't want to take the time to go through that process four times.

[Michael Stelzner]

Just out of curiosity, how much does it cost typically to create an image?

[Jonathan Mast]

About 3 and a half cents, give or take a little bit. So every image is going to cost you three to five cents.

[Michael Stelzner]

Okay, well, the idea of just creating four of them makes a lot of logical sense, I would imagine, because you're not going to love every single one of them, right?

[Jonathan Mast]

Exactly. I find that I toss at least 50% of the images that I've created because it's messed something up sometimes. As simple as messing up fingers. We know AI and imagery sometimes messes up fingers, so sometimes it's that. Sometimes it gives me the wrong type of beard, sometimes it just doesn't look like me for whatever reason. So just expect that's part of creating images with AI.

[Michael Stelzner]

Okay, There was something else you were about to say before I interrupted you.

[Jonathan Mast]

Yeah. No, and the last part is the output format. So you can choose to output as a. A JPEG file. A JPEG or a PNG file. Either will work just fine. I personally prefer a PNG because it's not a compressed format. And so if I need to bring that into Photoshop or something later, I've got more data for Photoshop to work with.

[Michael Stelzner]

Yeah. I also love PNG because it's considered lossless compression, meaning you won't lose any visual things when you compress it. And definitely the file sizes can be smaller on JPEGs for sure. So, I mean, I would imagine if you start with a png, is one more expensive than the other or are they all kind of the same?

[Jonathan Mast]

Nope, they cost the same. Yeah, it's just whether it compresses it or not. So why not get the highest quality I can at that point?

[Michael Stelzner]

Okay, so how long does it take to generate these four images, typically?

[Jonathan Mast]

So I do four images, it's probably going to take about 30 to 60 seconds for it to generate that on most days. Sometimes a little bit less, sometimes a little bit more, depending on how busy their servers are, but typically 30 to 60 seconds.

[Michael Stelzner]

Okay, so let's assume most of the listening audience is this the first time. How many variations of this are they going to be realistically experimenting with, because my guess is it's not going to nail it on the first render, right?

[Jonathan Mast]

It will not. It's one of the reasons we have Felix give you five different prompts, because some of the prompts are going to work out better than others. And again, I will run a prompt from Felix and if I like the output overall, I may run it again. It's important to understand, though, that AI images are not copy machines. In other words, let's say I run a prompt, Michael, and I love the image and I go, I want to create another one. It's going to be a different image the next time. It'll be similar, but it'll be different. They're not copy machines. So it's one of the reasons that if I find a prompt that I like, I may run that four or five times and get a series of again, 10 to 20 different images that I can use. Because I'll use my college campus one. I had me in a navy blazer, wearing jeans and a blue oxford and a hat, standing on a college campus. Sometimes I was standing in front of a building, sometimes I was standing in what looked more like a park that buildings in the background. Sometimes there were other students walking by. Sometimes I was in front of something that looked like, you know, the university sign out front. It will change those things as you're doing it. So you can't simply go, oh, I love this. This is the exact prompt and I can get another image just like it. It doesn't work that way.

[Michael Stelzner]

How about the resolution of the images? Are they generally decent resolution?

[Jonathan Mast]

They're very good. I use them on my website. I use them for social media. I don't think I try to put one on a billboard necessarily. But there are upscaling tools that you can upload these images to that can create much higher resolution images. So if you needed it for a print product or something like that, you can absolutely, what we call, upscale those images accordingly.

[Michael Stelzner]

So all the images that we generate, are they easy to locate in the future when we come back to this FAL AI tool, easy as a matter of perspective.

[Jonathan Mast]

But the answer is yes, in my opinion, they are so at the top when you go to the page that you mentioned. So if you put in like we talked about, socialmediaexaminer.com Laura that'll always bring you back to this page. You're going to see some tabs. One is called Playground. That's where we're at right now. Everything we've talked about is in the playground. You're then Going to also see one that's called Requests. And if you click on the Request tab, it will have all of your files that you've created and they're there to redownload. So if you don't download one right away and you want to go back, you just click on the Request tab and it's got all your history there.

[Michael Stelzner]

Okay. This is just a crazy idea, but what's your thoughts about eventually picking. Let's say we've eventually rendered dozens of these things out of the AI and we found some that are really good. What's your thoughts about training up an entirely new Flux Lora? Have you done something like this on the best AI generated images? So you increase the likelihood you get it really, really dialed in?

[Jonathan Mast]

I actually did. So we recently did that on my account and I found it made a bit of a difference. I'm not going to say it was a dramatic difference.

[Michael Stelzner]

Okay.

[Jonathan Mast]

But yeah, we took about 24 AI generated images and I went, I really like these. We uploaded it and I think it is probably more accurate than the first one I did, but not so noticeable that I would say, oh, this is something you've got to do. I don't. I don't know. It was worth the time and effort that I put into it.

[Michael Stelzner]

Talk to me a little bit about your thoughts on disclosure when it comes to utilizing these images to represent yourself.

[Jonathan Mast]

My opinion is if you're representing something that's not accurate, then you want to make sure that you're disclosing that. I'm. As a matter of course. Of course. I talk about AI, so people are not surprised when I'm using it, like on social media, but on my website, which I have a disclaimer on the website that lets people know that we do use AI to help us generate content that's on this website. We want to be very forthcoming about that. Now, I don't wear a neon sign, so to speak, around every image and goes, this image is generated by AI because the reality is, for example, me standing at a university campus in what appears to be Arkansas isn't misleading any of my customers. It's not misleading anybody in my audience. That's not something that's not true. Now, if I did a picture of me receiving the Presidential Medal of Honor and I was purporting that that was true, obviously that would not have integrity and it's not something I would recommend doing.

[Michael Stelzner]

Unless that's on your vision board. Right.

[Jonathan Mast]

That's a different spot. Yeah, if it's on your vision board. That'd be a great spot for it. Exactly.

[Michael Stelzner]

What about other creative applications of this? You mentioned cartoons and stuff like that. Can you use this? Because I know you do a lot of stuff that's like, cartoony looking. Do you create Flux Loras for that too?

[Jonathan Mast]

Yeah, I do. I actually have a style that we talked about that's trained in that. So I can now take, you know, Michael, I could take an image of you, for example, and I could upload that and tell it to render that in the style of a Pixar image. And my FAL implementation would do that for me. Certainly. If you're. If you're creating images of other people, though, then you also want to be very careful because you're using their likeness at that point in time, and that opens up a whole nother can of worms, shall we say, when it comes to fair use and everything else. But if you're doing it on your own, I think as long as you're not representing something that's, again, clearly not accurate, a vision board excluding from that, then I don't see any issues with it myself.

[Michael Stelzner]

Okay, I'm going to just ask a crazy question that we did not have prepared, but I've seen models where you can take images and you can slightly animate them or make videos out of them. Have you experimented with any of these? And what's your thoughts on taking some of your Flux Loras and using them inside of these?

[Jonathan Mast]

I have. In fact, it's a great way to have a little bit of fun and create some B roll video or background video, maybe transition video for something you're doing. And the nice thing is fal, the site that we're recommending, it hosts a number of those video models as well. So I can then take that image that I've got and I can upload it. And generally for about 10 cents per second of video, I can basically upload a picture again, give it a prompt, and say, I want, you know, a man with long white beard walking down the sidewalk at a university. Upload a picture of me standing at that university, and it will then literally animate that so that it appears that I'm walking down that sidewalk instead of just standing on the sidewalk. Most of them do a very good job, but again, much like with images, it's a bit hit or miss. So some are going to do exactly what you want, some are going to be off. I saw a good example in my Facebook group the other day, was a beautiful woman sitting on a sofa, and they'd instructed her to get up and walk to the kitchen, which you could see was in the background of the photo. And her standing up looked realistic. Her walking to the kitchen looked realistic, except one thing. She walked right through the sofa. And that's the type of thing that can happen with AI Video. And then you have to try again.

[Michael Stelzner]

Yeah, that's so funny. Jonathan Mast. This has been really eye opening and I cannot wait to see the images that everybody makes. If people want to connect with you on the socials, where do you want to send them? And if they want to learn more about your Facebook group and the different products and services you have, where also do you want to send them?

[Jonathan Mast]

Same place. If they just remember my name, it's Jonathan Mast. M A S T and go to jonathan mast.com linktree and I've got tons of free resources, how to contact me and everything else right there.

[Michael Stelzner]

Yeah. And folks, I would love it if you tag me and Jonathan on Facebook or whatever platform if you end up making some images. I know we're both on Facebook. It's be really cool to kind of see the output that you make as a result of this. And I do strongly recommend you check out his Facebook group. What's the name of your Facebook group? Because I know it's hopping.

[Jonathan Mast]

AI prompts for entrepreneurs. We got about almost 360,000 of us in there now.

[Michael Stelzner]

Jonathan, thank you so much for sharing your insights with us today.

[Jonathan Mast]

Thank you, Michael.

[Michael Stelzner]

Hey, if you missed anything, we took all the notes for you over@social mediaexaminer.com a48. And also be sure to follow this show on your favorite podcasting app. And if you've been a regular listener, would you do me a favor and give us a review and maybe let your friends know about this show? And do check out our other shows, the Social Media Marketing Podcast and the Social Media Marketing talk show. This brings us to the end of the AI Explored podcast. I'm your host, Michael Stelzner. I'll be back with you next week. I hope you make the best out of your day and may AI help you to become more successful. Successful. The AI Explored podcast is a production of Social Media Examiner.