AI portraits look real at thumbnail size 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 whenever they’re in frame. This workflow names each failure mode, gives you the exact prompt vocabulary that counters it, points you at the right model for June 2026, and walks the generate-and-retouch loop that finally produces a portrait you can hand a client.
TL;DR
- Three failure points, three specific fixes: prompt for skin texture (pores, not “beautiful skin”), eye catchlights (named light position), and hand pose (hide them or pose them simply at the side).
- Pick the model for the job: Midjourney v7 for cinematic editorial faces, Flux 1.1 Pro for skin micro-texture and commercial licensing clarity, Nano Banana 2 (Gemini 3.1 Flash Image) for fast iteration, SDXL/SD 3.5 locally when you need LoRAs and no per-image cost.
- Budget 6 to 10 generations per usable portrait, not 1 to 2 — realistic faces reject far more than stylized work.
- Always finish with a 5-minute retouch in Photoshop or Affinity. AI does ~90%; the last 10% is what reads as studio quality.
What actually breaks, and why
A portrait that survives editorial scrutiny lives or dies on three details the eye checks subconsciously in under a second:
| Failure | What it looks like | Root cause |
|---|---|---|
| Wax skin | No pores, uniform tone, plastic sheen | Models over-smooth toward an idealized “beauty” average |
| Dead eyes | No catchlight, flat iris, doll stare | The model wasn’t told where the light source is |
| Broken hands | Extra fingers, fused knuckles, melted joints | Hands have huge pose/anatomy variance and little training signal |
Fix all three or none — one wax detail collapses the whole illusion. The sections below give the prompt vocabulary for each.
Who this helps
Editorial and brand designers needing portrait imagery on a budget, indie creators building character references, founders generating press-kit profile photos, and freelancers producing avatars for client landing pages.
This is not the tool for a real person’s verified likeness (book a photoshoot — no model locks identity reliably), for stylized or illustrated portraits (use art-style prompts instead), or for anything legally requiring real identity such as ID photos or KYC headshots.
Which model to use (June 2026)
There is no single “best” model — they trade off realism, control, and cost. Picks below reflect the June 2026 landscape; defaults shift every few weeks, so re-test on your own prompt before committing a project.
| Model | Best for | Access / price (as of June 2026) | Character consistency |
|---|---|---|---|
| Midjourney v7 | Cinematic editorial faces, default-pretty lighting | Basic $10 / Standard $30 / Pro $60 / Mega $120 per month (annual ~20% off); no free tier | Omni Reference (--oref), ~90% across scenes |
| Flux 1.1 Pro / Flux.1 Kontext Max | Skin micro-texture, multi-subject scenes, commercial clarity | API ~$0.003–0.05 per image via providers | Kontext multi-reference, up to 10 images |
| Nano Banana 2 (Gemini 3.1 Flash Image) | Fast iteration, sharp detail, in-Gemini editing | API ~$0.045 (512px) to $0.151 (4K) per image; Pro tier in Google AI Pro $19.99/mo | Strong reference-image editing |
| SDXL / Stable Diffusion 3.5 | Full local control, LoRAs, zero per-image cost | Free, runs on your own GPU | LoRA / identity weighting (most setup) |
For pure facial realism in June 2026, Midjourney v7 and Flux 1.1 Pro lead — both render skin texture, hair detail, and eye reflections close to studio photography. For consistent characters across many shots, Midjourney’s Omni Reference and Flux’s Kontext multi-reference are the two strongest no-fine-tuning options.
Before you generate
- Decide the final use first: print headshot (300 DPI — generate at ~4x final size), web profile (1:1, 1024px), or editorial hero (3:4 at the largest size the model supports).
- Pull 1 to 2 reference portraits in the look you want — a magazine cover, a stock shot, a film still. “In the style of this reference” beats prose every time.
- Specify the subject: ethnicity, age range, body type, expression. Vague prompts default to the model’s average training subject.
- Plan a higher reject rate than stylized work: budget 6 to 10 generations per keeper.
Step by step
- Skin — prompt texture explicitly.
Natural skin texture, visible pores at cheek and forehead, subtle imperfections, soft shadow under jawline, slight unevenness in tone.This counters the model’s over-smooth-into-airbrush default. - Eyes — name the catchlight.
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.Catchlights are the single biggest realism test. - Hands — avoid them near the face. When you must show hands:
natural relaxed hand pose, four fingers and thumb clearly visible, knuckles correctly placed, no extra digits, hand at side or in pocket.Hands stay AI’s weak point and produce the most rejects. - Hair — add imperfection.
Slightly windswept with individual flyaway strands at temple, subtle variation in hair direction at hairline, natural shine without halo.Counters the “perfectly styled” advertising-shot default. - Lighting — one believable setup.
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 — speak photographer.
85mm portrait lens, f/1.8, slight bokeh, eye-level framing.Lens vocabulary is the most underused realism signal. - Generate 6 to 8 variants per prompt. Roughly 1 keeper in 8 is the realistic baseline for faces.
- Retouch the winner in Photoshop or Affinity: 5 minutes of skin-tone balance, stray-hair cleanup, and light dodge/burn on the face lifts the result above “obviously AI.”
Quality check before you ship
- Skin texture passes at 100% zoom — visible pores, faint shadow under the jaw, subtle color variation across the cheek. Airbrushed = fail.
- Both eyes have catchlights from one believable light source. Wax eyes read fake within half a second.
- Hands are either out of frame or anatomically correct. Count the fingers; check knuckle placement.
- Hair has natural flyaways and direction variation — no perfect helmet.
- Lighting is one believable setup, not two contradicting sources.
- Lens choice flatters the framing — 85mm or longer for head-and-shoulders; never 35mm for portraits.
When the face still melts
A profile headshot flow: prompt skin / eyes / hands / hair / lighting / lens → 8 generations → ~2 usable → light retouch → final export. When the face itself comes out warped despite good lighting and lens choices, run the distorted-faces quick-fix pipeline — adjusting framing distance, regenerating at higher resolution, and a face-restore pass usually rescues the gen.
Reuse the workflow
- Save the prompt with named slots (subject, lighting, lens, framing). For the next portrait, swap the subject only, generate, retouch.
- Keep 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 its reference image so you can recreate the look in six months.
- Re-test every 4 to 6 weeks. Model updates change defaults — your “natural skin texture” trick may stop being necessary.
Common mistakes
- “Beautiful skin” → airbrushed wax. Use “natural texture, pores, subtle imperfections.”
- Forgetting eye catchlights → wax-figure stare.
- Hands in awkward positions near the face → hide them or pose them simply at the side.
- Treating one bad output as failure → realistic faces need more attempts than stylized work.
- 35mm lens at head-and-shoulders distance → distorts the nose; use 85–135mm.
- Skipping retouch → AI does 90%; the final 10% of human work is what makes it commercially usable.
Advanced tips
- For older subjects, prompt age-appropriate texture:
fine lines around eyes, silver at temples, natural skin laxity at neck. Models default to youthful. - For ethnic diversity, name the ethnicity directly and add specifics (
warm brown skin, deep-set eyes, broad nose). Vague prompts collapse to one demographic. - Match light to skin tone — soft warm key for deep skin, slightly cooler key for fair skin.
- For two people in one shot, expect double the reject rate; couples and groups multiply identity and pose errors.
FAQ
- Why do AI portraits still fail on hands?: Hands carry huge anatomy and pose variance with relatively little clean training signal, so models still misplace fingers and knuckles. It improves each release but isn’t solved. When in doubt, crop or hide hands, or pose them simply at the side.
- Can I use AI portraits commercially?: It depends on the platform’s terms and your jurisdiction. Midjourney’s paid plans grant commercial use (Pro and Mega add Stealth mode for client confidentiality); Flux and SDXL outputs are generally commercial-friendly, but check the specific license. Never generate real public figures or trademarked likenesses, and disclose AI generation for ad use where policy requires it.
- What’s the best model for realistic portraits in June 2026?: For facial realism, Midjourney v7 and Flux 1.1 Pro lead; Nano Banana 2 (Gemini 3.1 Flash Image) is fastest for iteration, and SDXL / SD 3.5 give full local control with LoRAs. Test 2 to 3 on the same prompt and pick by output, not marketing.
- Do I always need to retouch?: For commercial, client, or press use, yes. Five minutes of skin-tone balance and stray-hair cleanup is what separates “AI looking” from studio quality.
- How do I keep the same face across many generations?: Use a character-consistency feature — Midjourney v7’s Omni Reference (
--oref, holds ~90% consistency across scenes), Flux’s Kontext multi-reference (up to 10 images), or a trained LoRA on SDXL. Without one, the face drifts.