How to Create Aesthetic AI Generated Pictures of Yourself With Prompts

The short version: aesthetic AI pictures of yourself need a trained personal model built from 15 to 25 of your own photos, not a single clever prompt typed into Midjourney. The prompt matters, but it's doing maybe 30 percent of the work; the other 70 percent is the quality of the training photos you feed the tool and the negative prompts you use to stop it from smoothing your face into someone else's.

I spent an evening in October last year trying to build myself a new LinkedIn photo using Astria, because my last professional headshot was five years old and I looked like I'd aged backwards according to my own memory of it. Twenty two selfies, one afternoon, and a training bill of nine pounds later, I had 200 images of a woman who looked like me on a very good day, in a blazer I don't own, in lighting my flat has never had. Some of them were usable. Most were not. That gap between "usable" and "not" is what this post is about.

What "AI pictures of yourself" means (there are two very different methods)

People use this phrase loosely and it causes confusion, so let's separate the two things named differently:

• Personal fine-tuning (LoRA or trained model): you upload 15 to 30 photos of your own face, the tool trains a small custom model on your features for 10 to 40 minutes, and then you can generate unlimited new images of "you" in any pose, outfit or setting by writing prompts. This is what Astria, Photo AI, and Leonardo AI's character reference feature do.
• Face reference or face swap prompting: you give the tool one photo and ask it to generate a new scene using that face as a reference, without any training step. Midjourney's --cref parameter and Google's tools work this way. Faster, cheaper, but less accurate to your actual bone structure.

If you want something that looks convincingly like you across dozens of images, use the first method. If you want one fun image for a party invite, the second is fine and takes five minutes.

Step one: build a training set that doesn't sabotage you

This is the part every tutorial rushes past and it's the part that decides whether your output looks like you or looks like a stranger wearing your haircut. From my own testing across three tools, here's what a good set needs:

• 18 to 25 photos minimum. Fewer than 15 and the model guesses too much; more than 30 and most tools start averaging your face into blandness.
• A genuine mix of angles: straight on, three quarter left, three quarter right, slightly from above, slightly from below. I skipped the "from below" shots the first time and every generated image gave me a longer neck than I have.
• Different lighting conditions. All-daylight photos train a model that can't cope with a warm indoor prompt.
• No sunglasses, no heavy filters, no group photos cropped down to your face. Cropped group shots are the single biggest cause of blurry or distorted AI portraits I've seen when reviewing other people's failed attempts.
• At least a few photos where you're not smiling. Every training set I've seen that's 100 percent smiling photos produces generated images with a slightly manic default expression, because that's all the model has learned.

Upload these to your chosen tool, let it train (Astria takes about 20 minutes, Leonardo's is closer to 10), and you get a model reference tag or trigger word, something like lilachwoman, that you insert into every future prompt.

Step two: the prompt formula that produces something aesthetic

"Aesthetic" is doing a lot of work in this topic and most people never define it. In practice, aesthetic AI portraits share a specific recipe: soft directional light, a shallow depth of field, a limited colour palette, and a camera reference that tells the model to behave like a real lens rather than a smartphone. Here's a prompt structure that has worked consistently for me across Midjourney and Leonardo:

[trigger word], [subject description], [outfit], [setting], [lighting], [camera and lens], [mood word], [style reference]

A working example from my own generations:

"lilachwoman, confident woman in her fifties with a silver bob, wearing a forest green wool blazer, standing by a large window in a minimalist office, soft natural window light from the left, shot on a Canon 5D with an 85mm lens, shallow depth of field, warm and approachable expression, editorial headshot photography style, muted colour palette"

Compare that to what most people type: "professional headshot of a woman." That vague version gives you generic stock-photo energy every time, because the model has nothing specific to hold onto. Specificity in the prompt is what separates an aesthetic image from a forgettable one.

Negative prompts matter as much as the prompt itself

This is the bit that gets left out of most guides. Without a negative prompt, you'll fight recurring problems: waxy over-smoothed skin, extra or fused fingers if hands are in frame, asymmetric earrings, and teeth that look slightly too white and too even. My standard negative prompt across tools is roughly:

"blurry, distorted hands, extra fingers, plastic skin, overexposed, asymmetrical face, unrealistic teeth, watermark, text, logo"

Adding "plastic skin" alone cut the number of unusable images in my last batch by what I'd estimate at 40 percent, just from a rough count of the 50 images I generated with it versus without it the week before.

Aspect ratio and platform: match the output to where it will live

Don't generate square images and then crop them badly for LinkedIn. Set the ratio at generation time:

• LinkedIn profile photo: 1:1, but leave headroom because LinkedIn crops circularly.
• LinkedIn banner: 4:1, and put your face on the right third, not centred.
• Instagram feed post: 4:5 gives you the most vertical real estate in the feed.
• Website "about" page hero: 16:9 or wider, with your face positioned off-centre so text can sit beside it.

In Midjourney you add --ar 4:5 to the end of the prompt. In Leonardo and most others it's a dropdown before you generate.

Which tool to pick

I've tested enough of these now to have real preferences, and cost matters more than most reviews admit:

• Astria: around £8 to £10 per trained model, then roughly 8 to 10 pence per generated image. Best face accuracy of everything I've tried, worst interface.
• Leonardo AI: free tier gives 150 tokens a day, enough for 20 to 30 images; paid plans start around $12 a month. Good middle ground for beginners.
• Midjourney: $10 a month for the basic plan, no personal model training built in without third-party workarounds, but unbeatable for pure aesthetic quality on backgrounds and lighting.
• Photo AI: built specifically for "AI photos of yourself," one-off packs from around $29, very beginner friendly, less flexible on prompt control.

If you're deciding between these for anything beyond personal use, especially for a business or client-facing brand, I've written a longer comparison covering accuracy, licensing and cost per image in how to compare AI image generators for business use, which goes into the commercial licensing traps that most "best AI image generator" listicles skip entirely.

The uncomfortable part nobody selling you these tools wants to say out loud

Here's the bit that most posts on this topic avoid because it's bad for engagement: these models are trained on datasets that skew heavily toward youth, thinness and conventional symmetry, and left unprompted they will quietly shave years and pounds off you. I'm 53. Without specific instruction, every tool I tested returned a version of me that looked mid to late 30s, smoother skin, fewer lines around the eyes, a slightly different jaw. That's not a compliment, it's a training bias, and if you don't correct for it in the prompt you'll end up with an "aesthetic" photo that nobody who has met you in person would recognise. My fix was blunt: I now add "woman in her fifties, visible fine lines, natural skin texture, no retouching" directly into the prompt every single time, and it works, though you have to fight the model's default a little on every generation. If you skip this step, don't be surprised when your AI headshot gets a comment along the lines of "is this even you."

There's a second uncomfortable truth here too: using a heavily idealised AI photo as your professional headshot without disclosing it is a small trust risk. I've had a handful of new contacts mention, gently, that I looked "different" in person after meeting from my LinkedIn photo. It didn't end anything, but it stuck with me enough that I now pick the generations closest to reality rather than the most flattering ones.

A realistic workflow, start to finish

If you want to do this this weekend, here's the actual sequence rather than a vague suggestion to "experiment":

• Gather 20 to 25 photos of yourself today, on your phone, over one sitting if needed, in different rooms for different light.
• Pick one tool (Leonardo if you want free, Astria if you want accuracy).
• Train the model. Wait the 15 to 20 minutes. Don't skip through the setup screens, the crop tool matters.
• Write your first prompt using the formula above, generate a batch of 4, review, adjust the negative prompt, generate again.
• Expect to generate 40 to 60 images to get 4 or 5 usable ones. That ratio is normal, not a sign you did it wrong.
• Download at the highest resolution offered, then run it through a basic sharpening pass if your platform compresses images heavily (LinkedIn does).

That whole process, training included, takes about 90 minutes the first time and closer to 20 minutes every time after, once your model exists and you're just writing new prompts against it.

A quick note if you're doing this instead of paying for a real photographer

An AI headshot session costs you an evening and under £10. A proper photographer for a set of professional headshots in the UK typically runs £150 to £400 for a couple of hours and 10 to 15 edited images. The AI route is cheaper and faster, and for social media and casual use it's fine. For anything that matters, a Forbes feature, a speaking bio, a book jacket, I'd still book the photographer, because the fine detail work, the real texture in fabric and skin, still isn't quite there yet on close inspection. Use AI for volume and speed, use a real camera for the one image that has to hold up under scrutiny.

And if you're generating a big batch late at night waiting for renders to finish, which happens more than you'd think because most tools queue you during busy periods, it's not the worst use of a spare ten minutes to have something offline lined up rather than doomscrolling; I keep a list of free games to play offline on my phone for exactly that kind of dead time, and there's a separate list of games that work with zero data if you're somewhere with a bad signal.

I keep every related walkthrough in the ChatGPT Prompts and Limits: 30 Guides to Fun Prompts, Photo Edits and Daily Caps. To write your own, try the free tool that writes the prompt for you.

Frequently asked questions

How many photos do I need to train an AI model of myself?

Between 18 and 25 works best across most tools. Under 15 produces guesswork and inaccurate features; over 30 tends to average your face out and reduce how much it looks like you.

Why do my AI photos of myself look younger and thinner than I am?

The underlying models are trained on datasets that skew toward youth and thinness, so without specific correction in your prompt they'll default that way. Add explicit descriptors like your actual age range and phrases such as "natural skin texture, no retouching" to counter it.

What's the cheapest way to make aesthetic AI pictures of yourself?

Leonardo AI's free tier gives you 150 tokens a day, enough for 20 to 30 image generations, and if you already have a decent set of your own photos you can start producing usable results without spending anything.

Can I use AI generated photos of myself as a professional headshot on LinkedIn?

Yes, plenty of people do, but pick the images closest to how you look rather than the most flattering version, since an overly idealised photo can cause an awkward gap when people meet you in person after connecting online.

Useful references

• Anthropic's own documentation
• the NIST AI Risk Management Framework