Choose photos for the intended result
The best reference depends on what you are creating:
- AI family portrait — separate people into one photoreal group (including multi-generation).
- Add person to photo — place someone (often memorial) into an existing scene.
- A clear single-face source for animation after restoring the still, because motion makes blur and facial errors more visible.
A BringBack studio family portrait or add-person edit uses 2 credits. If the face is very small, blurred, covered, or shown at an extreme angle, choose a better source or rescan before generating.
Why the source image limits likeness
When you upload a reference for AI family portrait generation or adding someone into a photo, the model needs visible facial structure: eyes, nose bridge, mouth corners, jawline, and brows. Face-alignment research illustrates why those structures matter (for example, 300-W / 68-point schemes). When the source does not contain enough usable detail, generated features can depart from the person in the reference, a failure known as identity drift.
Practical check: aim for a face at least 200 pixels tall in the source. Smaller faces may work, but the risk of identity drift rises quickly as real detail disappears. If possible, rescan the original or choose a closer portrait first (see scan family photos safely).
Rule 1: Camera Angle — Front or Three-Quarter View
The AI needs to see both eyes, the nose bridge, and both mouth corners in the same frame. This is only possible in two angle ranges:
Front-facing (0° yaw)
Both eyes are equally visible. The nose appears centered. This gives the AI maximum data and produces the most accurate likeness.
Three-quarter view (15–45° yaw)
One ear may be hidden, but both eyes and the full nose bridge are still visible. This is the most common angle in casual portraits and works well. Beyond 45°, one eye starts to disappear behind the nose bridge, and the AI must hallucinate the hidden side.
Angles to avoid
Full profile (90°) — only one eye visible. The AI has to invent the entire hidden half of the face. The result will look like a plausible face, but it won't be their face.
Extreme upward or downward tilt (>30° pitch) — distorts the apparent distance between eyes and mouth. The AI may reconstruct a face with proportions that feel “off” because it learned from mostly level photos.
Rule 2: Face Resolution — The 200px Minimum
The face in your reference photo needs to be large enough for the AI to detect individual features — not just “there is a face here” but the specific shape of this person's iris, nose tip, and lip contour.
| Face Width | What the AI Sees | Likeness Quality |
|---|---|---|
| 400px+ | Individual pores, iris texture, eyelash direction | Excellent |
| 200–400px | Eye shape, nose profile, mouth contour, skin texture | Good |
| 80–200px | General face shape and eye position, but features are blurred | Usable |
| Under 80px | A blob with eye-socket shadows — AI will generate features | High drift risk |
How to check: Open the photo on your computer, zoom to 100%, and look at the face. Can you clearly see the shape of each individual eye? Can you distinguish the upper lip from the lower lip? If yes, the resolution is sufficient. If the face is a soft blur of skin tone, scan or photograph the source again at higher resolution. See our scanning guide for DPI recommendations.
Rule 3: Lighting — Even and Soft
Lighting directly affects how the AI interprets facial geometry. The model reads brightness transitions to understand where the nose protrudes, where the cheekbones sit, and where the jawline ends. Bad lighting creates false signals.
What works
Soft, diffused light from a window or overcast sky. Even studio lighting with minimal shadows. Indoor photos taken without flash. The face should be evenly illuminated with no more than a gentle gradient from one side to the other.
What causes problems
Hard direct sunlight — creates deep, sharp-edged shadows under the nose and brow that the AI may interpret as actual facial contours. A harsh noon shadow under the nose can make the AI think the upper lip is much thicker than it actually is.
Side-lit dramatic portraits — half the face is in deep shadow. The AI cannot see the landmarks on the dark side and must hallucinate them, which increases drift risk.
Direct camera flash — flattens the face and blows out skin texture. The AI loses the subtle depth cues (cheekbone shadows, nose bridge gradient) that help it distinguish one person from another.
Rule 4: Clear Face — No Obstructions Over Key Landmarks
The AI prioritizes five landmark regions. If any of these are covered, it will generate that region from its training data instead of from your actual photo:
- Eyes — the most identity-carrying feature. Sunglasses, heavy shadows, or hair strands across the eyes force the model to guess eye shape, color, and gaze direction.
- Eyebrows — shape and thickness are highly individual. Wide-brim hats or heavy bangs that cover the brows lose critical identity data.
- Nose bridge and tip — defines the central axis of the face. A hand resting against the cheek or a scarf pulled up to the nose removes this anchor point.
- Mouth and lip contour — especially the Cupid's bow shape and commissure (corner) angles. Avoid photos where the person is mid-bite, mid-yawn, or holding something in their teeth.
- Jawline — defines face shape (oval, square, heart). Turtlenecks, scarves, or a hand cupping the chin obscure this boundary.
Glasses are usually fine. Standard prescription glasses with clear lenses do not obstruct eye landmarks. The AI can see through them. Dark sunglasses or reflective lenses are the problem — they hide the iris and eye shape entirely.
Rule 5: Neutral or Gentle Expression
Extreme expressions distort facial geometry in ways that confuse AI landmark mapping:
- Wide-open laughing — stretches the mouth, squints the eyes, and raises the cheeks. The AI may reconstruct a face that looks permanently “squished.”
- Squinting in bright sun — narrows the eyes so much that iris shape is lost. The AI guesses a generic eye shape.
- Crying or extreme emotion — redness, tear streaks, and contorted features add noise that the model may interpret as skin texture or facial contours.
The ideal expression is a relaxed neutral face or a gentle closed-mouth smile. This keeps all 68 landmark points in their natural resting positions, giving the AI the most accurate map of the person's actual facial structure.
Checklists by outcome (so you do not waste credits)
Memorial composite / person who passed
- One clear face photo of the deceased (not a tiny crop from a crowd).
- Restore scratched/faded prints before merge.
- Prefer neutral expression; avoid funeral candids with extreme emotion if you have alternatives.
- Path: restore → portrait or add person (2 credits per result).
Multi-generation group
- Each person gets their own best face source (do not crop five faces from one blurry picnic).
- Match approximate life stage when possible (do not force “everyone at 25” from mixed decades unless intentional).
- Kids: use recent clear photos; avoid heavy filters and stickers.
Animate after restore
- Restore first; animation amplifies scratches and blur.
- Need a single clear face, front-ish, eyes visible (animation is 10 credits).
- See subtle vs exaggerated animation for old family photos.
Mixing Vintage Prints and Modern Phone Photos
Family portrait composites often combine a 1960s B&W print with a 2024 smartphone selfie. The AI handles the lighting and color matching, but it cannot fix fundamental problems in the source material. Here is the workflow:
Restore damaged vintage prints first
If the old photo has scratches, tears, or fading across the face, run it through old photo restoration before using it as a portrait reference. The portrait model needs clean landmarks — it is not designed to repair damage.
Match the era if possible
If you want a portrait of the family “as they were” in the 1970s, use photos from that era for everyone — not a mix of 1970s prints and 2024 selfies. Age differences in the same person across decades will confuse the AI's age estimation.
Use the highest-quality version available
If you have both a wallet-sized print and the original 4×6” version, always scan the larger one. More pixels = more facial data = better likeness. A 4×6” print scanned at 600 DPI gives you a 2400×3600px image — far more data than a phone snapshot of a wallet photo.
Quick Reference
| Factor | Use | Avoid |
|---|---|---|
| Angle | Front-facing or three-quarter (0–45°) | Full profile, extreme up/down tilt |
| Face size | 200px+ across the face | Tiny crops from group photos (<80px) |
| Lighting | Soft diffused, window light, overcast | Hard sun, half-face shadow, direct flash |
| Face visibility | All 5 landmarks clear (eyes, brows, nose, mouth, jaw) | Sunglasses, hand on face, hair across eyes |
| Expression | Neutral or gentle smile | Wide laugh, squinting, crying |
| Source quality | Original camera file or 600 DPI scan | Screenshot of a screenshot, social media thumbnail |
FAQ
Can I use a group photo as a source for one person?
Only if their face is large and sharp when cropped (ideally 200px+). Otherwise find a dedicated portrait.
Should I upscale a tiny face before uploading?
Upscaling cannot invent true identity detail. Prefer a better original or rescan. Blind upscale often adds plastic texture that still drifts.
How many credits if likeness fails?
Each studio portrait attempt uses 2 credits. If repeated results miss the likeness, change the reference photo rather than repeating the same input.