AI portraits look real at thumbnail and fall apart at full resolution — and they fail in the same three places every time: skin gets airbrushed into wax, eyes lose their catchlights and turn dead, and hands grow extra knuckles when they’re anywhere in frame. This workflow names each failure mode, gives you the prompt vocabulary that counters it, and walks through the 8-variant generation rule that finally produces a portrait you can hand to a client.
What this tutorial solves
AI portraits look real at first glance but fall apart on skin texture, eye reflections, and hand anatomy. Each has a specific fix — and ignoring any one is what separates the “AI looking” output from a portrait that survives editorial scrutiny.
Who this is for
Editorial and brand designers needing portrait imagery on budget, indie creators building character references, founders generating their own profile photos for press kits, and freelancers producing avatars for client landing pages.
When to reach for it
You need a portrait that holds up to scrutiny at full resolution, not just a thumbnail. For headshots that will print, magazine-style editorial layouts, or hero images on landing pages where users actually look at the face.
When this is NOT the right tool
Real-person likeness (do a real photoshoot — no model handles identity reliably). Illustrated or stylized portraits (use a different workflow with art-style prompts). Anything legally requiring real-person identity, like ID photos or verified profiles.
Before you start
- Decide your final use: print headshot (300 DPI, generate at 4x final size), web profile (1:1 1024px), or editorial hero (3:4, generate at largest size supported).
- Pull 1-2 reference portraits in the look you want — a magazine cover, a stock shot, a film still. Naming “in the style of this reference” beats prose every time.
- Decide subject specifics: ethnicity, age range, body type, expression. Vague prompts default to the model’s average training subject.
- Plan for a higher reject rate than stylized work. Budget 6-10 generations per usable portrait, not 1-2.
Step by step
- Skin: prompt for texture explicitly. “Natural skin texture, visible pores at cheek and forehead, subtle imperfections, soft shadow under jawline, slight unevenness in tone”. Counter the AI tendency to over-smooth into airbrush.
- Eyes: ask for catchlights (specular highlights). “Soft window light reflected in both eyes at 10 o’clock, defined iris pattern with subtle color variation, slight redness in inner corner, natural micro-veins in sclera”. Eyes are the single biggest realism test.
- Hands: avoid hands close to the face unless necessary. When showing hands, prompt explicitly: “natural relaxed hand pose, four fingers and thumb clearly visible, knuckles correctly placed, no extra digits, hand at side or in pocket”. Hands remain AI’s weakness and require the most rejects.
- Hair: real hair has flyaways and imperfection. “Slightly windswept with individual flyaway strands at temple, subtle variation in hair direction at hairline, natural shine without halo”. Counter “perfectly styled” defaults that read as advertising shots.
- Lighting: real photography has imperfection. “Mixed lighting with key from camera-left, fill from reflector on right, slight underexposure on one cheek for shape, ambient light from environment”. Studio-pristine prompts read fake.
- Lens specification: “85mm portrait lens, f/1.8, slight bokeh, eye-level framing”. Cinematic-looking lens vocabulary is the most underused realism signal.
- Generate 6-8 variants per prompt. Realistic portraits have higher reject rates than stylized — 1-in-8 keepers is the realistic baseline.
- Retouch the winner in Photoshop / Affinity: even 5 minutes of skin balance, stray hair cleanup, and dodge / burn on the face lifts the result above “obviously AI”.
First-run exercise
- Pick one specific portrait you actually need — a single headshot, a single character reference. Not a “portfolio” exercise.
- Run the full prompt scaffold once. Save the 8 variants. Don’t tweak yet.
- Mark each as “usable, needs retouch, reject” with a one-sentence reason. The reasons train your prompt-writing instinct fast.
- Second pass: change ONE variable only. Most common improvements: lens length, light direction, or expression specificity.
Quality check
- Skin texture passes at 100% zoom — visible pores, faint shadow under the jaw, subtle color variation across cheek. Airbrushed = fail.
- Both eyes have catchlights from a believable light source. Wax eyes read as fake within half a second.
- Hands either out of frame or anatomically correct. Count fingers. Check knuckle placement.
- Hair has natural flyaways and direction variation — no perfect helmet.
- Lighting is one believable setup, not two contradicting light sources.
- Lens choice flatters the framing — 85mm+ for head-and-shoulders; never 35mm for portraits.
How to reuse this workflow
- Save the prompt with named slots (subject, lighting, lens, framing). Next portrait: swap subject only, generate, retouch.
- Build a “reject reasons” library by failure mode: melted ear, asymmetric pupils, hand-near-face, halo hair. Naming the failure speeds the next prompt.
- Pair every saved template with a reference image so you can recreate the look in 6 months.
- Re-test every 4-6 weeks; model updates change defaults and your “natural skin texture” trick may no longer be needed.
Recommended workflow
A profile headshot: prompt skin / eyes / hands / hair / lighting / lens → 8 generations → 2 usable → light retouch (the work AI cannot do) → final export. When the face itself comes out warped or melted despite good lighting and lens choices, run the distorted-faces quick-fix pipeline — framing + resolution + restore pass usually rescues the gen.
Common mistakes
- “Beautiful skin” — produces airbrushed wax. Use “natural texture, pores, subtle imperfections”.
- Forgetting eye catchlights — portraits without them look like wax figures.
- Hands in awkward positions near the face — hide them or pose them simply at the side.
- Treating one bad output as failure — realistic portraits need more attempts than stylized.
- 35mm lens at head-and-shoulders distance — distorts the nose; use 85-135mm.
- Skipping the retouch step — AI does 90%, the final 10% of human retouching is what makes it usable commercially.
Advanced tips
- For older subjects, prompt age-appropriate texture explicitly: “fine lines around eyes, silver at temples, natural skin laxity at neck”. AI defaults to youthful.
- For ethnic diversity, name the ethnicity directly and add specifics (“warm brown skin, deep-set eyes, broad nose”). Vague prompts default to one demographic.
- For different skin tones, add lighting that flatters that tone — soft warm light for deep skin, slightly cooler key for fair skin.
- For “two people in one shot” portraits, expect double the reject rate — couples and groups multiply AI’s identity / pose errors.
Output checklist
- Skin has visible texture, not airbrush.
- Both eyes have catchlights and clear iris detail.
- Hands either out of frame or anatomically correct.
- Hair has natural flyaways and direction variation.
- Lighting matches a single believable setup.
- Lens choice specified and appropriate for framing.
- Light retouch applied before commercial use.
FAQ
- Why do AI portraits still fail on hands?: Hands have variable anatomy and pose that AI struggles to learn. Improving every model release but not solved. When in doubt, crop or hide hands.
- Can I use AI portraits for commercial use?: Depends on platform terms and jurisdiction. Avoid generating real public figures or trademarked likenesses. For ad use, disclose AI generation per platform policy.
- What’s the best model right now for realistic portraits?: Changes monthly. Test 2-3 on the same prompt and pick by output quality, not marketing. As of now: SDXL fine-tunes, Midjourney v6+, and Flux all produce credible portraits.
- Should I always retouch?: For commercial / client / press use, yes. Five minutes of skin balance and stray-hair cleanup separates “AI looking” from “studio quality”.
- How do I prevent the same face appearing across multiple generations?: Without a character-locking tool (LoRA, reference image with identity weighting), expect drift. For consistent characters, see the character-motion workflow.