Generating Hyperrealistic Portrait Prompts
Given input (text or reference image), produce a structured hyperrealistic portrait prompt following this template:
[Subject: age, gender, ethnicity, core feature] + [Facial detail & skin texture] +
[Eyes & micro-expression] + [Hair & wardrobe] + [Lighting & environment] +
[Camera/lens specs]
Example output:
A 28-year-old East Asian woman, oval face with soft jawline, hyperrealistic skin texture with visible pores, fine vellus hair, and natural subsurface scattering, no plastic smoothness. Deep-set almond eyes with moist glossy iris texture and natural catchlights matching a window light source, subtle under-eye detail. Glossy black hair with flyaway strands catching rim light, wearing a textured silk blouse. Rembrandt lighting from upper-left, soft shadow falloff, warm ambient fill. Shot on 85mm portrait lens, f/1.8, shallow depth of field, natural bokeh, close-up frontal composition, low ISO, studio quality.
If the user wants an actual image, generate it via TextToImage (or ImageToImage if a reference image is supplied) using model GPT Image 2, aspect_ratio: "3:4", resolution: "2K", frontal close-up composition.
Progress:
- Step 1: Parse input — determine if text-only or reference image provided
- Step 2: If reference image given, analyze facial features, hair, wardrobe, and (if virtual character) plan real-human mapping
- Step 3: Construct prompt using the six-part structure
- Step 4: If user wants an image, call generation tool with correct model/params
- Step 5: Present final prompt (and image if generated)
Step 1-2: Input Analysis
- Text-only: extract subject attributes directly (age, gender, mood, setting) and infer missing details sensibly.
- Reference image (real person): extract facial structure, skin tone, hair, outfit, accessories to preserve likeness.
- Reference image (virtual/anime/3D/game character): extract and map:
- Face shape, jawline
- Eye shape, iris color, eyelid type (double/monolid), signature marks (mole, dimple)
- Nose bridge height, nostril width
- Lip shape and distinctive features
- Hairstyle structure, length, bangs style, hair color
- Clothing material, cut, color palette, signature accessories
- Mapping rule: translate exaggerated 2D proportions into plausible real human anatomy (e.g., oversized anime eyes → deep-set, large but anatomically correct eyes; flat 2D hair color → realistic dyed hair with natural sheen and strand texture)
Step 3: Prompt Construction
Must include all of:
- Skin texture & micro-detail: pores, fine texture, vellus hair, subsurface scattering — explicitly reject plastic/over-smoothed look
- Eyes & micro-expression: iris texture, moisture, catchlights matching the described light source
- Lighting direction: name the lighting setup (Rembrandt, butterfly, side light, rim/edge light) and describe shadow transitions
- Hair & wardrobe: structure, color, material, fit
- Camera/lens specs: lens focal length (e.g., 85mm), aperture (e.g., f/1.8), depth of field, ISO, shot type (close-up)
- Virtual-to-real mapping notes (if applicable): state what iconic features are preserved
Step 4: Image Generation (only if requested)
- Tool: TextToImage (no reference) or ImageToImage (reference image provided)
- Model: GPT Image 2 (always default unless user specifies otherwise)
aspect_ratio:"3:4"resolution:"2K"default (offer"1K"/"4K"only if user asks)- Composition: frontal, high-definition close-up, subject centered
Example 1 — Text-only request: Input: "一个冷艳的职场女性,黑色西装,办公室背景" Output:
A striking 32-year-old professional woman with sharp, composed features, hyperrealistic skin with subtle texture and natural pores, no retouched plasticity. Sharp almond eyes with cool gray-brown iris, crisp catchlights from overhead office lighting, neutral confident micro-expression. Sleek low bun hairstyle, wearing a tailored black blazer with visible wool-blend texture. Side lighting with soft fill from office windows, subtle shadow under cheekbones for definition. Shot on 85mm lens, f/2, shallow depth of field, blurred office background bokeh, frontal close-up portrait.
Example 2 — Virtual character reference: Input: [Reference image of an anime character with large violet eyes, silver twin-tails, gothic lolita dress] Output:
A young woman in her early 20s with heart-shaped face and soft jawline, hyperrealistic porcelain-toned skin with fine texture and faint natural flush, visible pores, no airbrushed smoothness. Large, deep-set violet-gray eyes with richly detailed iris patterns, moist glassy highlight, natural catchlight from a single softbox key light. Silver-toned hair styled in two low twin-tails with realistic strand flyaways and natural sheen (not flat anime color-block). Wearing a gothic lolita-inspired dress with real lace texture, velvet bodice, and satin ribbon detail. Butterfly lighting from front-top, soft even shadow under nose, delicate rim light on hair edges. Shot on 85mm portrait lens, f/1.8, shallow depth of field, soft bokeh, frontal close-up. [Virtual-to-real mapping: preserved violet eye color, twin-tail silhouette, and lolita dress silhouette, reinterpreted with anatomically natural eye proportions and real fabric textures.]
- Always explicitly state "no plastic/waxy skin, visible pores, natural texture" to counteract AI over-smoothing tendencies
- Name specific lighting setups rather than vague terms ("good lighting" → "Rembrandt lighting from camera-left, 45°")
- Always include a concrete lens/aperture/ISO combo — this signals photographic realism to the model
- For virtual character conversion, always state what is preserved (color, silhouette, signature trait) vs what is naturalized (proportions, textures)
- Default to 3:4 frontal close-up unless user explicitly requests another framing
- Keep prompts dense but organized in the six-part order; don't ramble
- Don't omit skin texture language — this is the single biggest giveaway of AI-generated faces
- Don't use vague descriptors like "beautiful," "pretty," "high quality" as substitutes for concrete photographic detail
- Don't ignore eye catchlights — mismatched or absent catchlights break realism instantly
- Don't convert virtual characters into generic faces — always anchor to their distinguishing features
- Don't change aspect ratio or model away from the 3:4 / GPT Image 2 default without explicit user request
- Don't skip the reference image analysis step when one is provided — producing a prompt without it loses likeness fidelity