A consistent AI character is recognizable as the same person across portraits, product ads, UGC videos, locations, camera angles, and wardrobe-safe variations. The face matters, but facial similarity alone is not enough. Hairline, apparent age, body proportions, skin, hands, core clothing, and footwear can all drift when the model receives only one attractive portrait.
OpenCake Actors solve this by creating one private three-view character sheet with GPT Image 2 and reusing it as the identity reference. The sheet combines a large portrait with faceless front and rear body views, giving downstream models one canonical description of who the person is. This page explains the design and publishes the production prompt.
Why character consistency fails
A portrait strongly describes a face but weakly describes height, shoulder-to-hip relationship, torso-to-leg ratio, rear hairstyle, garment construction, and footwear. When later scenes require a full-body or rear view, the model must invent those missing details. Repeating prose can help, but prose does not carry the same visual specificity as a reviewed reference.
Identity also drifts when references contain conflicting hairstyles, different apparent ages, beauty filters, changing body shapes, or removable accessories. A useful character sheet reduces those contradictions before campaign production begins.
The OpenCake three-view character sheet
| Region | Purpose | Framing |
|---|---|---|
| Left portrait, about 47% | Locks facial identity, skin, hairline, and expression baseline | Eye-level head and shoulders with the complete hair silhouette |
| Center body, about 28% | Locks frontal anatomy, proportions, core wardrobe, hands, and footwear | Straight-on front view from immediately below the chin through the soles |
| Right body, about 25% | Locks rear hair, anatomy, garment construction, and footwear | Straight-on rear view with no face, profile, cheek, or reflection |
The body panels deliberately exclude the face. This gives the portrait enough space to preserve facial detail while the other panels document anatomy and wardrobe at a useful full-body scale. It also prevents the generator from creating three competing facial interpretations inside one reference.
Real OpenCake character sheet examples
The exact GPT Image 2 character sheet prompt
This is the production prompt used to turn approved front and rear actor references into the private identity sheet. Replace the character name and attach the two views. If only a written brief is available, first turn it into a specific character definition covering apparent adult age, observable face geometry, hair, skin, build, proportions, distinguishing features, core garments, and footwear.
Character name: <CHARACTER NAME>. Create one square 1:1 ultra-photorealistic studio character sheet from three separately photographed views of the exact same adult person shown in the two attached reference images, assembled into seamless vertical panels. These are not three people standing together.
The first reference is authoritative for frontal identity, face, anatomy, skin, hair, core garment front, and footwear. The second reference is authoritative for rear anatomy, rear hair, core garment construction, and footwear. References define identity and core wardrobe only. Every removable accessory visible in either reference is transient styling and MUST be removed: no hats, jewelry, earrings, necklaces, bracelets, rings, watches, bags, glasses, handheld objects, or removable hair accessories. This removal rule has higher priority than reference preservation. Reconstruct the person's natural unobstructed hair and head shape wherever a hat or accessory concealed them, consistent with visible hair, roots, length, texture, and the rear reference. Show bare ears, neck, wrists, and hands. Preserve the core garments and footwear, not the accessories. Treat the references as source material only; do not reproduce their separate canvases or add extra people.
Assemble three seamless vertical panels on one continuous pale gray-white studio sweep with no borders, gutters, labels, text, logos, or scenery. Left panel occupies approximately 47% of the canvas: a large eye-level head-and-shoulders portrait of this exact person, showing the complete hair silhouette, full face, neck, shoulders, and upper chest. Use a calm closed-mouth neutral expression, relaxed jaw, direct natural gaze, and low-distortion portrait perspective. Center panel occupies approximately 28%: the exact full-body straight-on front view, framed from immediately below the chin through the soles so the entire face is outside the frame and invisible. Right panel occupies approximately 25%: the exact full-body straight-on back view from the lower rear hair and upper neck through the soles, with no face, profile, cheek, or reflection visible. Keep both body figures at identical scale with complete footwear and comfortable edge clearance.
Preserve the exact identity, apparent age, ethnicity, skin tone, facial structure, eye shape and color, nose, lips, brows, hairline, hair color, hair length and texture, height, build, torso-to-leg ratio, shoulder-to-hip relationship, waist placement, limb proportions, hand scale, foot scale, core garments, garment colors, seams, hems, and footwear from the references. Lock these across all three views. Do not redesign, simplify, replace, homogenize, or restyle the character or core wardrobe. Make the person exceptionally photogenic and immediately memorable while remaining fully human; appeal comes from facial harmony, vitality, presence, and distinctive features, not artificial perfection or altered anatomy.
Render a real photographed person rather than a synthetic avatar. Preserve restrained left-right differences in the eyes, brows, nostrils, lips, jaw, ears, shoulders, waist, hands, knees, and stance. Use only three to five observed distinguishing imperfections; keep them incidental and never amplify or showcase them. Show spatially varied pores mainly around the nose and inner cheeks, fine vellus hair, faint expression lines, under-eye texture, mild localized redness, subtle tonal variation, natural joint color, realistic translucency at ears, nostril rims, and lips, individual hair strands, root variation, a little visible scalp through the part, baby hairs, and controlled flyaways. Texture responds to light, with softer detail in shadow and no uniformly sharp pores. Eyes have moist waterlines, slightly off-white sclera with faint capillaries, natural lash spacing, non-glowing irises, and consistent catchlights. Body skin includes fine follicle texture and sparse body hair, slightly darker elbows, knees, and knuckles, and natural joint creases. Hands follow natural finger-length hierarchy with correct thumb orientation, cuticles, nail beds, joints, and creases. Clothing obeys gravity with accurate material thickness, weave, seams, compression, folds, and contact with the body. Shoes make believable weight-bearing contact with the floor.
Use neutral full-frame studio reference photography with soft large light from camera-left, gentle frontal fill, subtle contact shadows, restrained contrast and saturation, moderate microcontrast, natural sharpness, and realistic dynamic range. Resolve skin, hair, hands, fabric, and footwear clearly without HDR halos or crunchy sharpening.
No CGI, 3D render, illustration, waxy or plastic skin, porcelain smoothness, beauty filter, skin blur, airbrushing, excessive symmetry, generic influencer face, enlarged eyes, tiny nose, overfilled lips, uniformly sharp pores, deliberately showcased imperfections, uniformly distributed fake freckles, glowing irises, mannequin pose, flexing, impossible waist, exaggerated curves, elongated limbs, changed ethnicity, changed age, changed body shape, changed core outfit, identity drift, face visible in either body panel, extra people, duplicate anatomy, identical fingers, fused or extra fingers, malformed hands, floating hair, painted-on clothing, incorrect seams, missing shoes, cropped feet, mirrors, reflections, props, hats, jewelry, earrings, necklaces, bracelets, rings, watches, bags, glasses, handheld objects, removable accessories, text, logos, labels, watermark, borders, gutters, dark panel artifacts, or background scenery.Why the accessory rule is unusually strict
Hats, jewelry, watches, bags, glasses, handheld objects, and removable hair accessories are styling, not identity. If those objects remain inside the canonical reference, later generators may reproduce them in every scene or use them to hide important anatomy. OpenCake therefore gives accessory removal higher priority than reference preservation.
Core garments and footwear are different. They help define body scale and provide a stable neutral baseline. Campaign wardrobe can change intentionally later, but the reference sheet should first document one coherent person rather than several styling concepts.
What should stay stable and what can change
| Identity lock | Creative variable |
|---|---|
| Face geometry, apparent age, and skin tone | Expression and performance |
| Hair identity and natural hairline | Intentional campaign styling |
| Build, proportions, hand scale, and foot scale | Pose, camera angle, and framing |
| Core anatomical relationship across views | Location, lighting, and art direction |
| Approved Actor asset | Product, hook, dialogue, and call to action |
How to use a character sheet after generation
Save the approved sheet as one reusable Actor and pass that Actor reference into every generation that needs the person. Do not replace the visual anchor with a fresh prose description for each scene. Prose should describe what the Actor does, where the scene happens, and what may change; the character sheet remains responsible for identity.
- Use the same Actor sheet for every storyboard and clip in one campaign.
- Assign the Actor and product separate reference roles so their features are not blended.
- Change one major creative variable at a time when testing hooks, locations, wardrobe, or delivery.
- Compare each output with the sheet before approving style, motion, dialogue, or polish.
- Do not promote a drifted output into a new reference; errors compound when weak generations become anchors.
Create or reuse a persistent AI character: Open the OpenCake Actor Library
From character sheet to video
The preview clips below demonstrate that each sheet is not just a display card. It becomes the identity reference for a moving Actor. Character consistency in video still depends on model capability, shot complexity, motion, duration, lighting, and how many other references compete for attention, so every clip requires review.
Character sheet quality checklist
- The portrait, front body, and rear body clearly represent the same adult identity.
- The entire hair silhouette appears in the portrait and footwear is complete in both body views.
- No face, profile, cheek, reflection, or partial facial feature appears in either body panel.
- Hair, build, proportions, skin, core garments, and footwear remain consistent across all views.
- Hats, jewelry, bags, watches, glasses, handheld objects, and removable accessories are absent.
- Hands, limbs, garment seams, body weight, and floor contact are physically plausible.
- Skin looks spatially varied and human rather than plastic, airbrushed, or uniformly textured.
- The background is one continuous pale studio sweep without borders, labels, props, or dark artifacts.
Common character consistency failures
| Failure | Why it matters | Response |
|---|---|---|
| Face appears in a body panel | It introduces another facial interpretation at a smaller scale | Regenerate with the below-chin and rear-no-face framing repeated explicitly |
| Accessory remains | The item can become permanently attached to the identity | Remove it and reconstruct unobstructed hair or anatomy |
| Body views use different scale | Proportions become ambiguous to downstream models | Match figure height and preserve complete footwear |
| Skin is overly perfect | The Actor reads as a synthetic avatar and loses memorable specificity | Request localized texture and restrained natural variation |
| Outfit changes between views | The sheet stops describing one coherent studio session | Lock core garment color, material, seams, hems, and footwear |
Consent and responsible character creation
Use fictional adults, licensed talent, or real adults who gave informed permission for the intended AI and commercial use. Do not create a reusable Actor from a celebrity, public figure, creator, employee, customer, or private person without appropriate authorization. Face, voice, territory, duration, channels, and revocation terms should be documented when a real person is involved.
Frequently asked questions
Can one portrait keep an AI character consistent?
A portrait can preserve facial likeness in similar shots, but it gives weak evidence about full-body proportions, rear hair, wardrobe construction, hands, and footwear. A reviewed multi-view sheet is a stronger general identity reference.
Why are the body panels faceless?
The large portrait carries facial identity at useful resolution. Removing faces from the smaller body panels reduces conflicting facial interpretations and gives more space to anatomy, proportions, garments, and footwear.
Does the same prompt work for illustrated characters?
The identity and layout principles transfer, but this production prompt is specifically written for photorealistic adult Actors. Illustrated, animated, animal, or stylized characters need a style-specific sheet prompt rather than instructions about photographic skin and studio capture.
The bottom line
Character consistency improves when identity is decided once and reused visually. A strong sheet separates face detail from body evidence, removes transient styling, locks realistic human proportions, and gives every downstream model the same canonical Actor. Build the identity first; direct the performance second.