The right-hand image is not a clean “after.” It is the better of two edits made from the left-hand baseline. It has more restrained facial rendering, stronger knit texture, a less abstract plant, and deeper environmental focus. It also changes plant geometry, window and background rendering, and fine facial details. That distinction is central: the test supports a directional improvement in photographic treatment, not exact identity or pixel preservation.
| Requested change | Observed result | Release decision |
|---|---|---|
| Less beauty polish | More local skin variation and less uniform smoothing | Direction accepted |
| Deeper phone-photo context | Plant, wall, window, and sofa read more clearly | Direction accepted |
| Preserve person and scene | Broad identity, pose, clothing, crop, sofa, and window remain; fine identity and scene details move | Exact preservation failed |
| Change treatment only | The generator re-rendered content as well as treatment | Do not call it a filter |
Diagnose the frame as five linked systems
“Make it more realistic” is not an actionable review note. Before revising, classify the visible problem. This prevents the common mistake of damaging a face while the actual failure lives in depth, lighting, material response, or behavior.
| System | Over-produced signal | Grounded alternative |
|---|---|---|
| Capture | Perfect exposure, portrait blur, extreme wide-angle proximity, cinematic grade | Plausible distance, ordinary exposure, scene-appropriate white balance, meaningful context |
| Face | Uniform surface, over-sharpened eyes, designed symmetry | Existing natural characteristics, restrained microtexture, local tonal variation |
| Hair | Helmet silhouette or thousands of decorative flyaways | Recognizable hair structure with irregular outer edges and a few individual strands |
| Materials and light | Painted-on clothing, floating objects, decorative reflections | Seams, folds, tension, contact shadows, and one coherent lighting system |
| Behavior and space | Centered presentation pose in a showroom | One small action, task-directed gaze, believable hands, ordinary surroundings |
These systems interact. Visible pores will not rescue a face lit like a cosmetics advertisement. An imperfect crop will not rescue a hand that does not contact the object it supposedly holds. A lived-in room will not help if every background edge is dissolved by synthetic bokeh. Ask which two or three signals dominate the artificial look, then revise only those.
Translate camera mythology into visible evidence
“Shot on iPhone,” “convert to RAW,” “one camera,” and “no portrait mode” sound concrete, but they mix style, hardware, file format, and software behavior. OpenAI's image prompting guidance treats camera language as high-level direction and warns that detailed specifications may be interpreted loosely. Apple describes ProRAW as a real DNG capture format that combines RAW information with iPhone processing. Apple describes Portrait mode as a depth-of-field effect that keeps the subject sharp while blurring the background.
A generated PNG can resemble an ordinary phone photograph. It does not thereby acquire a sensor, lens, device, capture mode, DNG payload, EXIF history, or documentary truth. Translate the desired appearance instead.
| Mythic instruction | Visible requirement | Boundary |
|---|---|---|
| Convert it to RAW | Neutral processing, restrained sharpening, believable dynamic range | Never claim a generated PNG is a sensor RAW or ProRAW file |
| Use one camera | One coherent viewpoint and perspective | Hardware count is not visible provenance |
| No portrait mode | No synthetic subject cutout; keep nearby context legible | Natural depth falloff can still exist |
| Standard wide smartphone | Eye-level view from a plausible human distance | Explicitly reject extreme wide-angle facial distortion |
| Automatic exposure | Ordinary highlights, shadows, and scene-appropriate color | Do not add ugly clipping merely to simulate a phone |
Do not manufacture a person to prove realism
The original checklist included asymmetry, blemishes, under-eye texture, shine, and other “imperfections.” Those cues can be useful for a new fictional character when handled gently, but they become risky in reference editing. A model should not invent a scar, skin condition, illness, fatigue, injury, age change, or facial difference that was not present in the source. “Healthy and ordinary” is also a poor instruction: it treats health and normality as an appearance category.
The safer rule is simple. For a new fictional character, ask for restrained natural variation without specifying a diagnosis or defect. For an existing reference, preserve the person's existing natural characteristics and change only the photographic treatment. Realism should emerge from capture, light, material, depth, and behavior—not from adding damage to a face.
Separate generation and reference editing
A universal prompt that tries to handle both jobs becomes long and internally tense. Generation asks the model to invent a coherent subject and scene. Reference editing asks it to protect existing identity and geometry while changing a narrow property. The copyable prompt page therefore makes the short generation version primary and provides a separate edit version.
The final compact generation version was also filled once and executed twice on the same bookstore scenario. Both outputs followed the requested realism direction, yet they produced different fictional people, poses, lower-body wardrobe, shops, and lighting balance. That limited pair is published on the prompt page. It demonstrates why prompt reuse and character continuity are different problems; two runs are not enough to estimate repeatability.
- Choose one job. Generate an original fictional adult, or edit one permitted reference. Do not leave both branches active.
- Name one small action. Adjusting a bag strap, repairing a radio, or straining tea gives the gaze, fingers, elbows, clothing, and objects a shared reason to exist.
- Lock capture geometry. Use a plausible distance, eye-level viewpoint, and no extreme wide-angle distortion. State whether nearby context must stay legible.
- Retain natural characteristics. Ask for subtle microtexture and tonal variation without inventing blemishes, asymmetry, fatigue, or age.
- Make materials carry the light. Name two or three important materials and their folds, seams, wear, reflection, or contact shadows.
- Exclude the wrong genre. Remove beauty campaign, studio portrait, influencer pose, fake bokeh, and cinematic grade only when those are the observed failures.
- Score the whole frame. Reject attractive images with broken hands, drifting identity, illegible generated text, inconsistent reflections, or false provenance.
The exact constrained edit prompt
The second edit restarted from the untouched baseline. It did not use the first drifting edit as a new source.
Open the exact executed reference-edit prompt
Use case: identity-preserve photo edit.
Input image: Image 1 is the only edit target.
Primary request: Change only the photographic treatment so the image resembles an ordinary, unretouched smartphone snapshot rather than a beauty portrait.
Visible treatment changes: reduce beauty retouching; retain natural existing skin characteristics with restrained pores and local tonal variation; keep natural lip texture, a mild forehead and nose sheen, a few individual flyaway hairs, realistic knit texture and folds, ordinary automatic exposure, neutral color, and slightly deeper environmental focus. Keep the result believable and subtle.
Hard invariants: preserve exactly the same fictional adult's identity, face shape, age, ethnicity, hair length and part, expression, gaze, seated posture, hand positions, navy sweater, light trousers, body proportions, crop, camera angle, window frame, sofa, cushions, plant geometry, background object positions, and image dimensions. Do not invent blemishes, scars, illness, injury, fatigue, age changes, makeup, props, or new background details. Do not change pose or composition. No synthetic portrait-mode blur, glamour lighting, cinematic grading, over-sharpening, fake film grain, text, logo, or watermark. Change only the rendering treatment; keep all geometry and content unchanged.The result is useful precisely because it is imperfect evidence. Compared with the first edit, it preserves the broad crop and pose much better. Yet a close inspection still finds a different plant arrangement, different background definition, and small facial-rendering changes. The correct status is “direction accepted, preservation partial,” not “same photo, fixed.”
Test outside one demographic and one lighting setup
Three similar young East Asian women would not meaningfully test the method. We added two new cases to the existing rainy bus-stop result. The small matrix now includes different visible age cues, skin tones, gender presentation, hair structure, actions, and outdoor, fluorescent-window, and warm-cool mixed lighting. This is better coverage, not demographic proof: every scenario still has only one generated candidate.
The radio, screwdriver, two hands, shirt, counter, pegboard, window, and fluorescent fixture form one readable working environment. Age texture is present without becoming the subject of the frame. There is also a concrete failure: the receipt contains marks that resemble unreadable generated writing. If legible paperwork matters, remove it from the generation brief, crop it, or make a dedicated text edit and verify every character.
This case tests a different failure surface: dense curls, two hands around a small action, reflective metal, patterned ceramic, a dark window, warm interior light, and cooler exterior light. The action and gaze align; the shirt hangs with weight; the room remains legible. One successful-looking candidate cannot establish consistent treatment of South Asian subjects, curly hair, hands, or mixed light. Repeated and independently reviewed trials would be required for that claim.
The twelve-question review matrix
Pass means the visible candidate met the narrow editorial check. Partial means usable direction with a material caveat. Fail means the relevant goal was not met. “Pass” never means real camera provenance, universal model capability, or production approval.
| Question | Baseline | Edit v2 | Bus stop | Repair shop | Kitchen | Book A | Book B |
|---|---|---|---|---|---|---|---|
| 1. Skin has restrained local variation | Fail | Pass | Pass | Pass | Pass | Pass | Pass |
| 2. Eyes and gaze agree with scene and action | Partial | Partial | Pass | Pass | Pass | Pass | Pass |
| 3. Lips and mouth edges avoid painted smoothness | Partial | Pass | Pass | Pass | Pass | Pass | Pass |
| 4. Facial variation is plausible, not manufactured | Partial | Pass | Pass | Pass | Pass | Pass | Pass |
| 5. Hair structure and outer edges are believable | Partial | Pass | Pass | Pass | Pass | Pass | Pass |
| 6. Expression belongs to the stated action | Fail | Fail | Pass | Pass | Pass | Pass | Pass |
| 7. Hands, grip, contact, and occlusion are coherent | Partial | Partial | Pass | Pass | Pass | Pass | Pass |
| 8. Clothing seams, weight, folds, and tension agree | Partial | Pass | Pass | Pass | Pass | Pass | Pass |
| 9. Light and reflections form one scene | Pass | Pass | Pass | Partial | Pass | Partial | Partial |
| 10. Depth keeps useful context without fake cutout | Fail | Pass | Pass | Pass | Pass | Pass | Pass |
| 11. Framing is plausible at the stated distance | Partial | Partial | Pass | Pass | Pass | Pass | Pass |
| 12. Synthetic provenance and release limits are explicit | Pass | Pass | Pass | Pass | Pass | Pass | Pass |
The matrix exposes three points a beauty-only review would miss. First, the apartment edit still fails the action-expression check because it remains a camera-facing portrait. Second, the repair-shop case is only partial on light because the bright fixture and window are close to clipping, and it has a separate text-artifact failure outside the twelve core questions. Third, both bookstore runs contain book-spine marks that should not be treated as verified text, and their bright fixtures or windows are close to clipping. The public record contains per-run notes, dimensions, hashes, prompts, and the superseded edit failure.
Log failed edits instead of polishing the story
The first reference edit increased context and surface detail, but it also changed the face, crop, hands, plant, window view, cushion, and surrounding layout. It was visually plausible yet unsuitable as a controlled comparison. We did not erase it from the evidence. We classified it as a failed edit, restarted from the untouched baseline, tightened the invariants, and made one second attempt.
The second attempt still did not satisfy every invariant, so it remains a partial result. This failure log is more useful than a claim that the prompt “removes AI look.” It tells a practitioner when the method stops: if protected identity or scene details matter, visual improvement cannot compensate for drift.
Use the method in an AI drama or commercial pipeline
A believable lead character can carry more value than a complicated plot only when the character survives production. This prompt helps establish the photographic treatment of one approved frame. It does not solve character continuity, performance, animation, lip sync, rights, or campaign results.
- Approve an original adult character. Record identity, face proportions, hair, age cues, wardrobe, and prohibited changes.
- Approve one realism reference. Use the five-layer review to select a treatment, not merely the prettiest face.
- Freeze continuity separately. Test front, three-quarter, profile, full-body, expressions, and lighting with the character-consistency workflow.
- Test motion separately. A strong still does not prove stable teeth, eyes, hands, hair, fabric, or face geometry across frames.
- Apply destination review. Confirm likeness permissions, disclosures, platform rules, claims, and human approval before commercial release.
Do not cite a single-image test as evidence of ad performance, audience trust, conversion lift, or a market price. Those questions require campaign data and a defined comparison—not a more photographic face.
Limits of this visual evidence
The narrow evidence is useful. On one undisclosed built-in image system, a structured prompt produced three standalone candidates with coherent actions, useful context, restrained texture, and believable material cues. A second reference-edit attempt moved treatment in the desired direction while preserving more of the source than the first attempt. Local checks verified the files, dimensions, color mode, hashes, and comparison render.
The evidence is not a benchmark. The earlier scenarios were run once; only the final filled bookstore prompt was run twice. The operator knew the prompt and scored the output. There was no independent reviewer, fixed seed, larger repeat set, real-photo ground truth, blind detection test, demographic study, animation test, audience study, sales study, or comparison with another generator. The built-in workflow did not expose the exact model identifier or revised internal prompt. Current results may not reproduce after a model update.
Frequently asked questions
Should I add pores, asymmetry, and blemishes?
For a new fictional person, restrained natural variation can help. For a reference edit, preserve what already exists; do not manufacture defects or health signals. If the portrait still feels artificial, check light, depth, action, materials, and camera distance before adding more facial detail.
Does “iPhone photo” help?
It can work as style shorthand. Follow it with visible constraints: plausible distance, ordinary exposure, meaningful context, restrained processing, and no synthetic portrait blur. Do not claim a real device or capture history.
Should I request RAW?
Not as a provenance claim. A real ProRAW file is a DNG camera capture with device processing. Ask for an unretouched visual treatment, then choose only output formats the tool actually supports.
Why not keep one long universal prompt?
Long prompts can work, but mixed generation and edit branches are harder to debug. A short base, one job, and optional scene modules reveal which clause changed the output.
Will this keep the same actor across an AI series?
No. It controls the appearance of one frame. Character continuity needs a reference board, trait inventory, multi-angle tests, and a separate release gate. Motion continuity needs another test again.
