Generic prompts like “professional portrait of a woman” produce the same mannequin every time: symmetrical, plastic-skinned, lit from straight on. This tutorial gives you a five-slot prompt scaffold and the lighting and lens vocabulary that gets headshots, character portraits, and marketing images that look photographed instead of rendered. Read this if you are tired of generating 30 images to find one usable one.
TL;DR
- Write portraits in five slots: subject + expression + wardrobe + lighting + lens. Vague style words (“beautiful”, “professional”) collapse every output to the model’s average face.
- Lens is the single biggest realism lever: 35mm distorts the nose, 85-135mm flatters. Always add texture words (“natural skin texture, visible pores, catchlight in both eyes”) because every 2026 model defaults to airbrushed.
- Tool pick (June 2026): Flux 2 Pro for top photoreal skin (~$0.03-0.05/image via API), Midjourney v7 for stylized character work (Basic $10 / Standard $30 / Pro $60 / Mega $120 per month), GPT Image 2 in ChatGPT for prompt-following and quick iteration (Free Instant mode; multi-image consistency on Plus $20/mo).
- Generate 6-8 variants per prompt, then iterate by changing one slot at a time. A “1 keeper out of 8” reject rate is normal; “1 out of 30” means the prompt is wrong, not the model.
What this covers
Realistic, flattering AI portraits built from five prompt slots (subject, expression, wardrobe, lighting, lens), the right tool for the look you want, plus the iteration discipline that keeps you from rewriting the prompt every time the face looks off.
Who this is for
Founders generating their own profile photo, creators making character references for fiction or games, indie marketers building avatars for landing pages, and anyone who keeps getting uncanny-valley results from default settings.
When to reach for it
LinkedIn, X, and About-page headshots; stylized character portraits for D&D or fiction; marketing avatars for testimonials (with consent); book-cover concepts; and dating-app or profile imagery where a real photoshoot is overkill.
Pick the right tool first
The model you choose moves the realism dial more than any single prompt word. As of June 2026, three tools cover almost every portrait need.
| Tool | Best for | Price (June 2026) | Character consistency | Notes |
|---|---|---|---|---|
| Flux 2 Pro | Photoreal skin, commercial headshots | ~$0.03-0.05 / image via API (fal, Replicate, Together) | Strong with reference image | 32B-param model; renders pores, subsurface scattering, correct hands ~95% of generations |
| Midjourney v7 | Stylized portraits, fantasy characters | Basic $10 / Standard $30 / Pro $60 / Mega $120 per month | Omni Reference tab (replaces old --cref, strength 300-500) | Standard+ adds unlimited Relax mode; best artistic output |
| GPT Image 2 (ChatGPT) | Fast iteration, prompt accuracy, text in image | Free (Instant); Plus $20/mo for Thinking mode | Persistent identity; up to 8 coherent images in Thinking mode | ~99% text-rendering accuracy, 4K output; runs on the GPT-5.4 image backbone |
Rule of thumb: if a client needs a photo that survives 100% zoom, generate on Flux 2 Pro. If you want a recognizable character across many scenes, use Midjourney v7 Omni Reference or GPT Image 2 Thinking mode. For a quick “good enough” headshot inside a tool you already pay for, ChatGPT is the lowest-friction option. New to Midjourney? Start with the Midjourney beginner guide.
Before you start
- Decide whether you want a real-photo look, a stylized look, or somewhere between. The prompt language differs sharply, and so does the best tool (Flux 2 Pro for photo, Midjourney v7 for stylized).
- Collect 2-3 reference images (a celebrity headshot, a stock photo, a painting) so you can name the look concretely instead of guessing. All three tools above accept a reference image.
- Pick aspect ratio up front: 1:1 for avatars, 3:4 for portraits, 9:16 for phone wallpapers. Switching mid-run wastes credits.
- Decide your reject threshold ahead of time. “1 keeper out of 8” is honest; “1 out of 30” means the prompt is wrong, not the model.
Step by step
- Subject slot. Start with the five-slot scaffold:
subject + age + expression + outfit + setting. Example: “woman, late 30s, soft smile with eyes engaged, charcoal blazer over white tee, soft-grey studio backdrop”. - Lighting slot. Add lighting in plain photography terms: “soft window light from camera-left”, “golden-hour rim light”, “beauty dish above and slightly right”. Avoid abstract words like “cinematic” alone; pair them with a concrete source.
- Lens slot. Add lens, aperture, and framing: “85mm, f/1.8, eye-level, shoulders-up”. The lens shapes face geometry: 35mm distorts noses, 85-135mm flatters.
- Realism slot. Add texture words last: “natural skin texture, visible pores, catchlight in both eyes, fine flyaway hair, sharp focus on iris”. Without these, every current model trends toward airbrushed.
- Batch. Generate 6-8 variants from the same prompt. Realistic portraits have higher reject rates than stylized art, so never judge from a single output. In ChatGPT, switch the model picker to Thinking to get up to 8 consistent variants in one pass.
- Iterate. Pick the best 1-2 and change ONE slot at a time: only the lighting, only the expression, only the lens. Mixed changes mask which variable helped.
First-run exercise
- Pick one persona you actually need (your own headshot, a single character), not a “portfolio” of ten different faces.
- Run the full five-slot prompt once and save the raw output. Do not tweak yet.
- Mark each output as “usable, needs retouch, or reject” and write one sentence explaining why for each. This builds vocabulary fast.
- For the second pass, change only ONE variable, most often lighting direction or lens length.
Quality check
- Does the face hold up at 100% zoom? Look for melted ears, mismatched earrings, asymmetric pupils, and floating hair strands.
- Are the eyes alive? Both irises should have catchlights from a believable light source. Wax-eyed portraits read as fake within half a second.
- Is the skin textured, not airbrushed? Pores, faint shadows under the jaw, and color variation across the cheek separate “photo” from “render”. Flux 2 Pro handles this best of the three tools.
- Does the outfit make sense for the lighting? A linen shirt in studio strobe looks wrong; a wool coat in golden hour also looks wrong.
How to reuse this workflow
- Save your winning prompt as a template named by purpose (“linkedin-headshot-male-40s”, “fantasy-character-rogue”) and only change the named variables next time.
- Build a small library of reject reasons: “wrong jawline”, “plastic skin”, “hand near face”, “iris asymmetric”. Naming the failure speeds the next prompt.
- Re-test your template every 6-8 weeks. Major model updates (GPT Image 2 shipped April 2026; Flux 2 and Midjourney v7 both landed in 2026) shift defaults, and your “natural skin texture” trick may no longer be needed.
- Keep your reference image set with each template so you can recreate the look later, or hand it to a teammate.
Recommended workflow
Subject + expression + lighting + lens, then first 8 variants, then mark keepers, then iterate one slot. If the first pass comes back with warped facial features, jump straight to the distorted-faces quick-fix pipeline instead of rewriting the prompt; most face issues are framing and resolution, not wording. Need the same face across many images? See consistent character images. For client or commercial use, route the winner through a light Photoshop retouch pass (skin balance, stray-hair cleanup), which is still faster than re-generating to perfection.
Common mistakes
- Plastic skin from missing texture words. Defaults are tuned for “beautiful”, which translates to airbrushed; always include “natural skin texture, visible pores”.
- Wrong lens for the framing. 35mm at head-and-shoulders distorts the nose; use 85mm or longer for flattering portraits.
- Vague lighting like “good lighting”. Name the source (“window light”, “softbox”) and the direction (“from camera-left”).
- Hands near the face. Even Flux 2 Pro (~95% correct hands) misses sometimes; crop hands out or move them below frame.
- Judging from one output. Realistic portraits demand 6-8 generations per prompt before you can fairly evaluate.
- Changing three variables at once. You will never know which fix actually worked.
- Using
--crefin Midjourney v7. It was deprecated; the v7 system ignores it. Use the Omni Reference tab (strength 300-500) instead.
FAQ
- Which AI tool makes the most realistic portraits in 2026?: For pure photorealism, Flux 2 Pro leads, with naturalistic pores, subsurface scattering, and correct hands in roughly 95% of generations, at about $0.03-0.05 per image via API. Midjourney v7 wins on stylized and fantasy portraits; GPT Image 2 in ChatGPT is the easiest for fast iteration and accurate prompt-following.
- Why does my portrait look like everyone else’s AI portrait?: You used default style words (“beautiful”, “professional”). Replace them with specific photographic language: lens (85mm), light source (“window light from camera-left”), and texture (“visible pores, catchlight in both eyes”).
- Can I generate the same person twice?: Yes, with a character-locking feature. Use Midjourney v7 Omni Reference (strength 300-500), GPT Image 2 Thinking mode (up to 8 coherent images, needs ChatGPT Plus at $20/mo), or a reference image on Flux 2 Pro. For consistency across many scenes, see consistent character images.
- My subject’s ethnicity keeps drifting. What fixes it?: Name the ethnicity directly in the subject slot and add specifics (“dark brown skin, deep-set eyes, broad nose”). Vague prompts default to the model’s average training subject.
- Is ChatGPT’s free image generator good enough for a headshot?: For a quick avatar, yes. GPT Image 2’s Instant mode is free for all users and follows prompts well. For multi-image consistency or self-verified output you need the Thinking mode that ships with ChatGPT Plus ($20/mo).
- What aspect ratio for LinkedIn?: Generate at 1:1 (1024x1024) and crop to LinkedIn’s required ratio in the platform. Generating at the exact final size sometimes truncates the head.