you keep asking "what tool".. that's not why this looks real the realism is written into the prompt, not the model let me show you the exact lines 👇

made the stills in GPT Image 2, the video in Seedance 2.5 at 1080p. what actually sells the realism has nothing to do with resolution.. it's pore-level skin texture, real freckle placement, a face that's slightly tired instead of camera-ready, and a camera that behaves like a real phone (shake, autofocus hunt, exposure drift). full process below 👇

step 1: i upload my own photo, GPT Image 2 does the rest use a photo where your face is clearly lit, front-facing, no sunglasses, no heavy shadow across one side.. the cleaner your source photo, the less the model has to guess Prompt: [Use my character reference image. Keep her exact face, facial proportions, eye shape, blue-green eye color, freckles, eyebrows, nose, lips, light brown hair color, hairline, skin tone, and overall identity. Do not drift to a different woman.] Ultra-photorealistic vertical iPhone front-camera selfie of the same adult woman from the reference image lying comfortably in bed in the morning, holding the phone at arm’s length slightly above eye level. FACE & IDENTITY: Preserve the exact woman from the reference. Fair warm skin with a subtle natural sun-kissed tone. Distinctive blue-green eyes with realistic iris detail. Soft freckles naturally scattered across her nose and upper cheeks. Natural medium-full eyebrows. Small straight nose matching the reference. Soft naturally shaped pink lips. Same facial proportions, jawline, cheek structure, eye spacing, and natural asymmetry. Beautiful but completely believable and human. No makeup or extremely minimal natural-looking makeup. HAIR: Keep her exact light-to-medium brown hair with warm golden tones from the reference. Medium-long length. Naturally loose and slightly wavy. Morning-bed texture with a few imperfect strands around her forehead and cheeks. Hair spread casually across a beige linen pillow. Individual flyaway hairs visible. No perfectly styled salon hair. SKIN REALISM . CRITICAL: Real unfiltered iPhone skin texture. Visible pores. Fine peach fuzz. Natural freckles preserved. Subtle under-eye shadows. Tiny variations in skin tone. Slight natural redness around the cheeks and nose. Very subtle morning skin sheen. Realistic lip texture. Individual eyebrow hairs and eyelashes. No skin smoothing. No beauty filter. No airbrushing. No waxy or plastic appearance. WARDROBE: Simple fitted black camisole-style tank top with thin straps and a normal modest neckline. Bare shoulders naturally visible. Relaxed everyday sleepwear appearance. Beige linen bedding loosely surrounding her. Comfortable, casual, Instagram-appropriate morning look. No jewelry, sunglasses, or outdoor accessories unless naturally appropriate for the scene. POSE: She is lying against soft pillows with her upper body slightly raised. One arm extends naturally toward the phone. The camera is held slightly above eye level. She looks directly into the front camera with a calm, sleepy, neutral expression. Relaxed shoulders. Natural posture. Not posing like a fashion model. FRAMING: Vertical 9:16. Authentic arm’s-length front-camera selfie. Framed approximately from mid-chest upward. Face remains the main focal point. Slight wide-angle distortion typical of an iPhone front camera. Very subtle perspective enlargement closer to the phone. Slightly imperfect crop and composition. Feels like a spontaneous morning photo rather than a professional portrait. ENVIRONMENT: Cozy Mediterranean-inspired bedroom. Soft beige linen pillows and sheets. Warm off-white walls. A subtle woven or macramé wall decoration softly blurred in the background. Small bedside table and minimal natural decor. Quiet, uncluttered atmosphere. LIGHTING: Bright natural morning window light entering from one side. Soft directional sunlight across her face. Warm highlights in her light brown hair. Gentle shadow on the opposite side of her face. Natural exposure with slightly brighter window-side highlights. No artificial studio lighting. No dramatic cinematic lighting. IPHONE LOOK: Real modern iPhone front-camera photograph. Natural computational photography. Slight smartphone sharpening but not excessive. Realistic highlight rolloff. Natural white balance. Small exposure imperfections. Subtle front-camera lens character. No artificial portrait-mode cutout. No professional camera bokeh. No cinematic color grading. No HDR-heavy appearance. MOOD: Quiet lazy morning. Effortlessly pretty. Relaxed. Personal. Natural. Unplanned. Feels like she woke up, picked up her phone, and casually took one selfie. NEGATIVE: different woman, changed identity, brunette dark hair, black hair, altered eye color, brown eyes, missing freckles, exaggerated freckles, changed facial proportions, heavy makeup, glamour makeup, perfect studio skin, plastic skin, waxy face, beauty filter, excessive smoothing, airbrushed skin, doll-like face, CGI, 3D render, overly sharpened image, professional fashion shoot, dramatic pose, lingerie, revealing clothing, provocative pose, exaggerated neckline, studio lighting, cinematic grading, fake depth of field, perfect symmetry, warped arms, malformed hands, extra fingers, duplicated limbs, distorted phone perspective, morphing face. 9:16 vertical, ultra-photorealistic, high detail, authentic iPhone selfie realism.

step 2: feed that same selfie back into GPT Image 2 with a new prompt.. this is what actually locks her face for the video one photo only shows one angle. the video needs her to turn her head, tilt, get close to the lens.. so you generate a multi-angle sheet from the same reference first Prompt: [Use the uploaded reference image as the ONLY identity reference. Keep the exact same adult woman in all six panels. Preserve her face shape, blue-green eyes, freckles, eyebrows, nose, lips, light brown hair, hairline, skin tone, and natural asymmetry. Do not drift to a different woman.] Create one clean 3×2 face character sheet with SIX head-and-shoulders portraits of the same woman on a plain light-grey studio background. Views: 1. Front 2. 3/4 left 3. Full left profile 4. Full right profile 5. 3/4 right 6. Slight low angle Expression: neutral and relaxed, mouth closed, eyes open. Identity: Fair warm skin, pale blue-green eyes, soft freckles across nose and cheeks, natural medium-full brows, light-to-medium warm brown hair with a soft center part, naturally straight to slightly wavy texture, soft pink lips, slim straight nose, soft oval face, delicate jawline. Skin realism: Visible pores, peach fuzz, natural texture, faint under-eye shadows, slight redness, subtle T-zone sheen. No smoothing, no beauty filter, no airbrush. Keep identical across all six views: eye color, freckles, brows, face shape, nose, lips, jawline, hair color, hair length, hairline, skin tone, age, and identity. Lighting: Even soft studio lighting in every panel. Style: Ultra-photorealistic, 85mm portrait look, natural skin, sharp facial detail, 4K. Negative: different women, identity drift, changed face, brown eyes, dark hair, missing freckles, heavy makeup, plastic skin, beauty filter, AI-perfect symmetry, different hairstyle per panel, CGI, cartoon.

step 3: attach both images to Seedance 2.5, set it to 1080p, vertical 9:16, 20 seconds, audio on image_1 = the six-panel sheet, primary face reference image_2 = the bedroom selfie, body, wardrobe and setting reference prompt:] SCENE CONTEXT 20-second vertical front-camera selfie video, filmed in ONE UNBROKEN TAKE. One adult woman, alone, lying in bed in a beige-linen bedroom in the morning, talking straight into her own phone's front camera. She tells the viewer she is not a real person. Nothing else happens. Nothing else is in frame. POINT OF VIEW LOCK THE CAMERA IS THE PHONE. The viewer sees exactly what her phone's front camera sees. The recording device does not exist as a visible object anywhere in this video. The phone is NEVER visible in frame. No phone in her hand, no phone against her cheek, no phone body, no camera lenses, no phone edge, no screen, no phone shadow on her skin or the pillow, no phone reflection in her eyes or in any surface. This is NOT a mirror selfie and there is no mirror in the room. The arm and hand holding the camera are BEHIND the lens at all times and never enter the frame. No forearm, no wrist, no fingers, no knuckles, no thumb, no arm crossing the lower frame edge. Her other hand rests down on the bedding, below frame, and never lifts toward her face and never holds anything. Her eye catch-lights are soft irregular window shapes only. No small bright rectangle in her pupils. FORMAT MODE SINGLE CONTINUOUS TAKE. ZERO CUTS. No hard cuts, no jump cuts, no insert shots, no angle changes, no transitions, no fades, no slow motion, no speed ramps. The camera never stops recording. All framing changes come from her off-frame arm moving the camera closer and further away. Real-time motion from frame one to the last frame. REFERENCES image_1 = BODY reference. Exact height, slim naturally toned build, long-leg proportions, shoulder width, waist, limb length, posture. image_2 = FACE reference. PRIMARY identity source. Lock her facial anatomy for the entire take. image_3 = ROOM and WARDROBE reference. Same beige-linen bedroom, same fitted black ribbed thin-strap tank top, same window light direction, same casual framing. Use the references for identity, room and wardrobe ONLY. Do not copy any phone, hand or device that appears inside a reference image. IDENTITY LOCK Fair warm skin with subtle sun-kissed warmth. Pale blue-green eyes with real iris fiber detail and wet catch-lights. Soft freckles across the nose bridge and upper cheeks. Natural medium-full light-brown brows. Soft oval face, delicate jawline, slim straight nose, naturally full muted-pink lips, faint natural under-eye shadow, subtle facial asymmetry. Light-to-medium warm brown hair with soft golden tones, medium-long, naturally tousled, slightly wavy, loose center part. Exact hairline, length, density, texture and color held throughout. Because this is one continuous take, her face, freckles, eye color, nose, lips, jawline, hair, skin tone and proportions physically cannot change at any point in the clip. MORNING STATE She woke up minutes ago and it is visible. Heavy upper eyelids. Faint puffiness and natural bluish shade under both eyes. Eyes glassy and wet. One soft pillow crease on the cheek resting against the pillow, unchanged for the whole take. Hair flattened on one side with volume on the other, several flyaways standing up in the backlight. Lips slightly dry with fine vertical texture. Faint sleep flush across cheeks and nose. Bare face, zero makeup, zero product. PERFORMANCE She is not confessing, she is winning, and she is enjoying it quietly. Her objective is to make the viewer admit they were fooled, and she changes tactic three times to get there. She opens generous, like she is letting you in on something, loose and sleepy and slightly amused. Around "You think I'm a real girl" she turns teasing, chin dipping, eyes on the lens. On "I'm not" the warmth switches off mid-breath and the delivery goes flat and matter-of-fact, and she stays flat while the camera comes in on her eyes and her skin, presenting evidence rather than performing feelings. On the last line the warmth comes back, but the words underneath it are cold, so it reads faintly predatory. Eye life throughout: continuous micro-saccades, the gaze arriving at the lens a fraction before her head does, thought readable in her eyes before each line, blink rate slow and heavy, one long dissociated blink after "None of this is real", catch-lights alive and wet in every frame. Her eyes drop just below the lens three times across the take and come back, checking her own preview image, without any screen being visible. Physical life: low center of gravity, everything economical, no illustrating gestures, no eyebrow acting. The closer the camera comes, the less she moves, so only the eyes and the thought carry those moments. Real involuntary business only: one swallow before "I'm not", one small sniff, hair falling across her face and being left there, a shoulder settling deeper into the mattress near the end. Her head and neck do the adjusting, never her hands. VOICE Paste verbatim: "A woman in her early twenties, neutral American accent, recording a casual vlog for herself in bed. Soft mid-range voice with light morning huskiness and a dry edge on the low notes, unhurried half-thrown-away delivery with audible breath, uneven pauses and slightly swallowed word endings, sitting very close to the mic so the low end is thick and intimate, warm and faintly amused, and going quieter rather than louder on the lines meant to unsettle you." One take, one voice, no re-take tone. She sounds like a person talking, not a person reading. Sentences do not all land at the same tempo. Real breath between phrases, a small catch before "I'm not", slight vocal fry on the ends of the quiet lines. Volume drops on "None of this is real" instead of rising. Her voice gets noticeably warmer, closer and thicker as the camera nears her face and thins out again as it pulls back, because the phone mic is moving with the lens. Perfect lip-sync on every syllable including the breaths. SPOKEN DIALOGUE, EXACTLY "No one is going to tell you how to make AI videos look this real. So I will. You think I'm a real girl, right? I'm not. Not my eyes. Not my skin. Not my voice. None of this is real. You've been watching me this whole time… and you couldn't tell." FIRST FRAME Frame one is already her face filling the vertical frame, front-facing, direct eye contact, mouth already open on the first word, mid-exhale of a small quiet laugh. Nothing but her, the pillow and the bedding in frame. No hand, no phone, no empty room, no silence before speech. The first word lands within 0.2 seconds. OPTICS Modern phone front camera, roughly 80 degree diagonal field of view. Camera distance changes only because her off-frame arm moves it, from 45 cm at full extension down to 10 cm at the closest point. Her face is the dominant object at every distance and the framing never gets wider than upper chest. Mild natural wide-angle distortion that increases visibly as the camera comes closer, slightly enlarging the nose and forehead at 10 to 15 cm. No telephoto compression, no creamy portrait bokeh, no lens flare, no digital zoom at any point. CAMERA CHOREOGRAPHY Never stabilized. Her elbow rests on the mattress, so the shake is heavy and low-frequency, not digital jitter: a slow drift one way, a small late wrist correction back, a tiny sag as the arm tires, a bump when her shoulder shifts. Framing is never centered and never level. The horizon sits 3 to 6 degrees off for most of the take. All movement is one continuous arm performance, seen only as camera movement: 0.0 to 4.0s. Camera at arm's length, roughly 45 cm, slightly above her eye level. Small settling drift then a correction at 0.5s. 4.0 to 7.5s. The camera moves in about 15 cm in one unhurried, slightly uneven approach. 7.5 to 10.0s. The camera rises and moves in to roughly 10 to 12 cm from her eye and holds there, visibly unsteady at that distance with tiny corrections. Her eye, brow and cheek fill most of the frame. 10.0 to 12.3s. The camera slides down her face to her cheek at roughly 12 cm while she turns the cheek into the window light. 12.3 to 14.5s. The camera continues down to mouth height, still close, chin and lips dominating the frame with her eyes partly visible at the top. 14.5 to 16.3s. The camera travels back out to full extension in one relaxed motion, the frame widening and settling. 16.3 to 20.0s. Arm's length again, back to the opening framing, small drift and one final 2 cm sag in the last second. Recording ends mid-frame with no reach for a button. Every move has the acceleration and deceleration of a human arm. No constant-speed travel, no gimbal float, no dolly feel, no crop-in. Natural rolling shutter on the faster moves, real motion blur, visible sensor noise in the shadow side of her face. AUTOFOCUS AND EXPOSURE Continuous autofocus reacting to the changing distance. It holds on her eyes at arm's length, hunts for 0.4 to 0.9 seconds on every distance change, overshoots slightly past sharp, then snaps back. The longest hunt is on the approach to her eye at 7.5s, breathing past sharp twice before locking on individual eyelashes and iris fibers. Auto-exposure dips as her face fills the frame at close range and lifts again as the camera pulls out, with the window highlights clipping on the white bedding for a few frames before settling. Focus and exposure change only because the camera moves. ACTION AND DIALOGUE TIMING 0.0 to 4.0s, arm's length. Already speaking on frame one, riding the tail of a small laugh, loose and warm, eyes flicking just below the lens at 1.4s and back. "No one is going to tell you how to make AI videos look this real. So I will." 4.0 to 6.2s, camera moving in. Teasing now, chin dipping slightly, a small closed-mouth smile. "You think I'm a real girl, right?" 6.2 to 7.5s, camera stops moving. One swallow, then the smile leaves her face and the delivery goes flat. "I'm not." 7.5 to 10.0s, camera at her eye, 10 to 12 cm. One eye fills most of the frame with freckles and cheek. Real pores, peach fuzz, wet lower lash line, one natural blink mid-line. Voice thick and close. "Not my eyes." 10.0 to 12.3s, camera sliding down to the cheek. She turns her cheek into the light and the sheen moves across the skin. Fine vellus hair, tiny skin-tone variation, faint redness. "Not my skin." 12.3 to 14.5s, camera at mouth height. Natural lip texture, a small dry catch on one word, visible breath on the word endings. "Not my voice." 14.5 to 16.3s, camera travelling back out. Completely expressionless while the frame widens. Volume drops. A strand of hair falls across her cheek and she leaves it. One long dissociated blink as the line ends. "None of this is real." 16.3 to 20.0s, arm's length, opening framing. Slow head tilt, eyes staying on the lens through the turn and arriving before the head does. Warmth returns to the face. After the last word, one slow blink, then a small restrained closed-mouth smile that arrives late. "You've been watching me this whole time… and you couldn't tell." PHYSICS Hair moves with real inertia and settles a beat late after every head move, strands falling across her face and staying there. The pillow compresses and recovers slowly under her cheek. Linen creases and shifts under her shoulder weight. Tank fabric folds naturally and never clings or morphs. The camera carries the real weight of a forearm and that weight increases visibly by the end. LIGHTING One condition for the entire take. Soft morning window light from screen-left, roughly 5000K, white bedding and wall bouncing gentle fill into the shadow side of her face. Warm edge on the left of her hair and jaw. One small blown highlight where the light clips her cheekbone. The light source never moves. No hard rectangular shadow anywhere on her face or the pillow, since nothing is between her and the window. REALISM Authentic modern phone front-camera rendering. Visible pores across nose and cheeks, fine forehead texture, peach fuzz on the cheek edge, natural freckles, tiny skin-tone variation, faint redness around nose and cheeks, natural under-eye coloration, subtle T-zone sheen, individual eyelashes and brow hairs, natural lip texture, flyaway hairs catching the window light. No beauty filter, no skin smoothing, no airbrush, no glamour retouch, no portrait-mode cutout, no cinematic grade, no studio look, no CGI feel. CONTENT LOCK One person only, start to finish. Nothing man-made in frame except the bedding, the pillow and her tank top. No second voice, no offscreen voice, no pet, no product, no cutaway, no b-roll, no exterior shot, no on-screen text, no captions, no subtitles, no logo, no watermark, no music, no sound design. Room tone and her voice only, plus faint fabric rustle when she shifts. NEGATIVE phone visible, phone in hand, phone against cheek, phone body, phone edge, phone camera lenses, camera module, phone screen, phone reflection, rectangular reflection in the eyes, device shadow on the face, mirror, mirror selfie, holding an object, hand in frame, fingers in frame, thumb in frame, forearm in frame, wrist in frame, arm crossing the lower frame edge, second person's hand, any cut, hard cut, jump cut, insert shot, angle change, camera switch, transition, fade, dissolve, digital zoom, crop-in, stabilized footage, gimbal smoothness, dolly move, constant-speed camera travel, random digital jitter, slow motion, speed ramps, different woman, identity drift, changing face, changing facial proportions, changing eye color, brown eyes, bright blue eyes, dark hair, changed hairline, changed hairstyle, missing freckles, exaggerated freckles, heavy makeup, false lashes, plastic skin, waxy skin, beauty filter, over-smoothed skin, AI-perfect symmetry, perfectly centered framing, robotic voice, text-to-speech cadence, presenter voice, energetic delivery, uniform sentence tempo, eyebrow acting, dead frozen eyes, unblinking stare, changing body proportions, oversexualized posing, lingerie, cleavage-focused framing, morphing face, flickering face, warped eyes, distorted lips, teeth artifacts, lip-sync drift, malformed hands, extra fingers, third hand, extra people, background music, captions, watermarks, cinematic lighting, teal-orange grade, oversaturation. OUTPUT 20 seconds, vertical 9:16, ultra-photorealistic phone front-camera POV, single continuous take, native synchronized dialogue, real-time motion.

Always a fan of realistic prompts. Just tried it. Great results. Thanks for sharing.

looks amazing

Realism is in the prompt nuances and camera angles, not just the engine. Unreal natural shot.

Amazinggg :)

🙏

wow

most people can tell this is AI. especially the ones you are selling to.

Nice and very realistic 🤩

It's so real

Except she DIDN'T tell us how to make the AI video 😆

This is an eye-opener opener

The prompt lines you listed are doing the work. Same thing happens in video: the realism comes from describing motion and skin detail, not just asking for 1080p.

This is the sort of prompt that can lead to unexpected creativity.

This is amazing. But what if one doesn't have access to seedance2.5

Work on improving the simulated voice. Mismatch of reverb to the space the model is in gives away a lot. There is also a distinct AI quality to the voice itself that distracts from realism and breaks the spell. Try different microphone style types as they have different filtering aspects that could add realism. The images are getting very good but the audio gives it all away.

@abxxai I learned the hard way that nailing the prompt is like 80% of the work. Makes everything else fall into place.

Nice share

Realism comes from the prompt, not the model.

this is why prompt structure actually matters fr

Isnt seedance 2.5 blocking the realistic people, or using grid bypassing the safety filters?

the acting is so bad, though. there is that uncanny AI like cadence in the voice. Without sound, it's amazing. You turn on the sound, it's instantly recognizable as AI. At least for me...

think realism comes from the model? Think again.

Exactly, the prompt is doing the heavy lifting.

thx!!

But it’s precisely such nuances that create the atmosphere! Incredible