AI Headshot Prompts: How to Generate Professional Photos with Midjourney, Flux, ChatGPT, and Stable Diffusion

You want a professional headshot without paying a photographer or a dedicated AI headshot service. You've got Midjourney, or Flux running locally, or a ChatGPT subscription with image generation. Can you get a usable professional headshot by prompting these tools directly?

Short answer: sometimes. The results depend heavily on which tool you use, how you prompt it, and what you mean by "usable." This guide covers what actually works, model by model, with specific prompts you can copy and test.

The Core Problem: Identity Preservation

Here's the thing every prompt guide glosses over. General-purpose image generators are excellent at creating a photo of a professional person. They're terrible at creating a photo of you.

Midjourney, DALL-E, and base Flux don't know what you look like. You can describe "30-year-old woman with brown hair" and get a photorealistic image of someone who matches that description. But it won't be your face, your bone structure, your specific features. It'll be a plausible stranger.

This is the fundamental limitation. A headshot that doesn't look like you isn't a headshot. It's stock photography.

The workarounds:

ChatGPT (GPT Image): Upload a selfie as a reference and ask it to generate a professional version. Results vary wildly. Sometimes it captures your likeness reasonably well. Sometimes it generates someone who vaguely resembles you from a distance.
Flux with a LoRA: Train a LoRA on 15-20 photos of yourself, then generate headshots using that model. This is the DIY path that actually produces identity-accurate results. It takes time, technical knowledge, and a GPU.
Midjourney with reference images: Use the --cref (character reference) parameter with your photo. Likeness preservation has improved significantly in V7 and V8, but it's still inconsistent.
Stable Diffusion with DreamBooth/LoRA: Same concept as Flux. Fine-tune a model on your face, then generate from it. Older approach, still works.

If you're not willing to do any of these, you'll get a generic professional photo that doesn't look like you. That might be fine for a placeholder, but it's not a real headshot.

Prompts by Model

ChatGPT (GPT-4o / GPT Image)

ChatGPT handles headshot requests better than most general tools because it can work with uploaded reference photos. The conversational interface also makes iteration easier.

Basic professional headshot:

Using the attached photo as an exact reference for my face, generate a professional headshot. Keep my exact facial features, skin tone, and hair. Professional studio lighting, neutral gray background, wearing a navy blue blazer over a white shirt. Natural, confident expression. Shot at 85mm focal length, shallow depth of field.

Multiple style variations:

Using the attached photo as reference, create four professional headshot variations:
1. Corporate (dark suit, gray background, serious expression)
2. Creative (casual shirt, warm-toned background, relaxed smile)
3. Speaker/author (dark top, blurred office background, engaged expression)
4. LinkedIn optimized (bright, clean, approachable, light background)
Keep my exact face in all four.

What works: ChatGPT understands context well. You can iterate with follow-up instructions ("make the lighting warmer," "try a different angle"). It handles text-heavy prompts better than any other tool.

What doesn't: Likeness consistency is unreliable. You might get a great result on the first try or spend 20 attempts getting something close. Each generation can drift from your actual appearance. Hands and ears sometimes render incorrectly if they're visible.

Flux (Local or via API)

Flux is currently the best open-source model for photorealistic images. For headshots specifically, it produces the most natural-looking skin, lighting, and fabric textures. But without a trained LoRA of your face, you'll get a stranger.

Without LoRA (generic professional headshot):

Professional corporate headshot photograph of a [age]-year-old [ethnicity] [gender] with [hair description], wearing a [outfit description]. Shot in a professional photography studio with Profoto B10 lighting, Rembrandt lighting pattern. 85mm f/1.8 lens, shallow depth of field. Neutral gray seamless paper background. Subject looking directly at camera with natural, confident expression. High-end corporate portrait style. 4K resolution, photorealistic.

With a trained LoRA (your face):

photo of ohwx man, professional corporate headshot, studio lighting, Profoto softbox key light, 85mm portrait lens, f/2.0, shallow depth of field, wearing navy suit and light blue dress shirt, neutral gray background, confident natural expression, high resolution

(Replace ohwx with whatever trigger token you used during LoRA training.)

Technical notes for Flux users:

Flux responds well to camera and lens specifications. Including focal length, aperture, and lighting equipment names produces more photorealistic results.
Resolution matters. Generate at 1024x1024 minimum.
CFG scale between 3.5 and 7 works best for photorealistic portraits. Higher values increase prompt adherence but can create an artificial look.
Negative prompts: cartoon, illustration, painting, drawing, anime, deformed, ugly, blurry, low quality, watermark

Midjourney (V7 / V8)

Midjourney produces visually stunning images, but it has a distinct "Midjourney look" that can make headshots feel slightly too polished, too perfect. It's an aesthetic optimization engine, not a reality engine.

Basic headshot:

/imagine professional corporate headshot photograph, [description of person], wearing navy blazer, white shirt, neutral studio background, Rembrandt lighting, 85mm lens, shallow depth of field, photorealistic --ar 1:1 --style raw --v 7

With character reference (your photo):

/imagine professional LinkedIn headshot, studio lighting, gray background, confident expression, wearing business professional attire, 85mm portrait lens --cref [your image URL] --cw 100 --ar 1:1 --style raw --v 7

Key parameters:

--style raw reduces Midjourney's tendency to over-stylize. Essential for headshots.
--cref with --cw 100 maximizes face similarity to your reference photo.
--ar 1:1 for square output (LinkedIn, most profile photos).
Add --no cartoon, illustration, painting to keep it photorealistic.

What works: Lighting and composition are consistently excellent. Midjourney rarely produces bad compositions.

What doesn't: Even with --style raw, results tend to look slightly idealized. Skin is too smooth, eyes too bright, symmetry too perfect. It's the "uncanny beauty" problem. People might notice it looks too good, which paradoxically makes it less trustworthy as a headshot.

Stable Diffusion (SDXL / SD 3.5)

Stable Diffusion is the most flexible option but has the steepest learning curve. It's also the oldest approach and has been partially superseded by Flux for photorealistic work.

Basic headshot (SDXL):

professional headshot photograph of a [description], studio portrait, Rembrandt lighting, 85mm f/2.0 lens, shallow depth of field, neutral gray background, wearing [outfit], natural expression, 8k uhd, photorealistic, detailed skin texture

Negative prompt: cartoon, anime, illustration, painting, drawing, deformed, ugly, blurry, low quality, oversaturated, plastic skin, doll-like

With DreamBooth/LoRA (your face): Same concept as Flux. Train on your photos, use your trigger token, generate variations. The ComfyUI or Automatic1111 workflow guides for this are widely available.

Honest assessment: For headshots specifically, Flux has overtaken Stable Diffusion in quality and ease. If you're starting fresh, use Flux. If you already have a Stable Diffusion setup and trained models, it still works fine.

The DIY Reality Check

Here's what nobody tells you about the DIY approach.

Time investment: Training a LoRA takes 20-30 minutes of GPU time plus an hour of prep (selecting photos, captioning, configuring). Generating and curating results adds another hour. Total: 2-3 hours for someone who knows what they're doing. 5-10 hours for a first-timer including setup and troubleshooting.

Hardware requirements: LoRA training needs a GPU with at least 8GB VRAM (12GB+ recommended). A consumer NVIDIA RTX 3060 or better handles it. If you don't have a capable GPU, you'll need to rent cloud compute, which adds cost.

Quality consistency: Even with a trained LoRA, maybe 1 in 10 generated images will be genuinely good. You'll produce dozens to find a few worth using.

The likeness problem persists: LoRAs trained on limited photo sets (5-10 images) produce inconsistent results. You need 15-20 diverse photos of yourself for reliable likeness. That's the same input a dedicated headshot service requires.

What you actually save: You save money (free if you have the hardware) but spend time and technical effort. For one headshot, the math rarely works out. For ongoing headshot generation (quarterly updates, multiple styles, team headshots), the investment pays off if you enjoy the process.

When DIY Works and When It Doesn't

DIY works well when:

You already have Flux/ComfyUI or Stable Diffusion running locally
You enjoy the technical process of training and generating
You want complete control over every aspect of the output
You need headshots frequently and don't want recurring costs
Privacy is a primary concern and you want everything processed locally

Use a dedicated headshot tool when:

You want results in minutes, not hours
You don't have a capable GPU or don't want to set up the toolchain
You need consistent, predictable quality without technical troubleshooting
You're generating headshots for a team and need a streamlined process
You value your time more than the $20-50 a dedicated service costs

Neither approach is wrong. It depends on what you value: control and customization, or speed and convenience.

For a deeper look at how dedicated AI headshot generators work compared to general-purpose tools, see our comparison of ChatGPT vs. dedicated headshot generators and our guide to how AI headshot technology actually works.

Frequently Asked Questions

Can ChatGPT make a professional headshot?

It can generate a professional-looking headshot, but likeness preservation is inconsistent. Upload a clear selfie and be specific about what you want. Expect to regenerate several times before getting something that looks like you. For a quick placeholder, it works. For a photo you'll use across professional platforms for months, the inconsistency is a problem.

What's the best AI model for headshots in 2026?

For photorealism: Flux (especially with a trained LoRA). For ease of use: ChatGPT with image upload. For aesthetic quality: Midjourney V8 with --style raw. For identity accuracy without technical setup: a dedicated headshot service that handles the fine-tuning for you.

Do I need to train a LoRA for AI headshots?

If you want the output to look like you specifically, yes. Without fine-tuning on your photos, every general-purpose model generates a plausible stranger, not you. The exception is ChatGPT, which can use an uploaded photo as reference, but its likeness preservation is less reliable than a trained model.

How many photos do I need to train a headshot LoRA?

15-20 photos work best. Include variety: different angles, lighting conditions, expressions, and distances. Avoid sunglasses, heavy filters, or group photos. The same requirements apply no matter if you're training your own model or uploading to a dedicated service. For more detail, see our source photo checklist.

Is it cheaper to DIY my headshot with Midjourney or Flux?

In dollar terms, usually yes. A Midjourney subscription costs $10-60/month and Flux is free to run locally. A dedicated headshot service costs $20-50 per session. But factor in your time: 2-5 hours of setup, training, and curation vs. 15 minutes of uploading and waiting. Your hourly rate determines which is actually cheaper.