Bad photo for AI projection: what gets in the way
Quick facts
- Shadow
- Hard shadow under the nose, lip or cheekbone is copied and can be read as relief.
- Angle
- A turned or tilted face changes the proportions the app uses to position the areas.
- Expression
- Smile, open mouth and tensed forehead stay in the projection; it does not change expression.
- Image
- Filters, portrait mode, blur and low resolution produce less faithful results.
AI projection seems like magic until the first bad photo. A slightly turned face, a shadow under the lip, a subtle smile, a phone filter: none of it looks serious to the naked eye, and all of it shows up again in the simulation. Harmy's AI does not fix the photo. It receives the image as it is and changes only what was marked, with the ml chosen. Everything else, including the flaws, comes along.
This page lists what gets in the way most, why, and how to spot the problem before spending a projection. The positive requirements for the photo are in one frontal photo for planning; here are the negative ones.
Why does the starting photo weigh so much?
Harmy generates an image from a frontal photo and the marked areas. The prompt forbids makeup, skin retouching, and changes to expression, background, hair and lighting. That is good: the projection is subtle and realistic, and the result does not "invent" a new face. The flip side is that it inherits everything the photo brought. Light, angle, expression and texture pass straight into the After.
In addition, generative models can make image errors, called hallucinations in image AI: a detail that did not exist or a strange texture. Generative AI is more stable with a clean photo. The worse the input, the greater the chance of an unexpected result.
Which problems degrade the simulation most?
| Problem | What happens | How to avoid it |
|---|---|---|
| Hard shadow | The shadow is copied and can look like relief or volume loss | Diffuse, frontal light (lighting for clinical photos) |
| Turned or tilted face | Proportions and the position of the areas are deformed | True frontal, Frankfurt plane |
| Smile or open mouth | The expression stays in the After and distorts lips and cheekbone | Neutral expression (neutral expression) |
| Beauty filter | Skin and contour arrive already altered | Camera without filter (filters and beautification) |
| Portrait mode or HDR | Blurred edges and altered texture | Normal camera (HDR and portrait mode) |
| Blurry photo | The AI invents texture | Locked focus (exposure and focus) |
| Close-up selfie | The face arrives distorted by the lens | Documentation distance (distance) |
| Partial or cropped face | Detection fails or the points land out of place | Whole face, with margin |
The beauty filter deserves a mention. It smooths the skin, slims the face and enlarges the eye, all before you take the photo. The projection starts from a face that is not the patient's, and the comparison with the actual result loses its meaning.
How does shadow mislead the projection?
Shadow is relief information. When the photo has a strong shadow under the cheek, the AI interprets it as part of the volume and may leave it as is, or accentuate it. The result is an After with contours that look deeper or fuller than intended. Light from above creates shadow under the nose and lip; side light creates shadow on one side of the face only. Diffuse light reduces the problem.
The ideal photo does not need a studio. A large window with soft light, the patient facing it and a plain background solve much of it. Watch out for the window as a light source, which can become side light if the patient is poorly positioned, and for shadow on the face.
And the angle?
The app detects the face and positions points based on facial structure. A turned or tilted face changes the distances between those points, and the marked area lands out of place. In the projection, the simulated filler may appear shifted or asymmetric because of the head position. A chin up or down also alters the reading of the chin, lips and nose.
Align the head to the Frankfurt plane, with the eyes horizontal, the gaze on the horizon (gaze on the horizon) and the face forward, with no neck rotation. Watch out for tilted face and symmetry.
How to spot a bad photo before spending the projection?
Each projection costs R$ 3.99, and a new account gets 1 as a welcome gift. It is worth spending a minute checking the photo before generating.
If any item fails, retake. If the app does not detect the face, the problem is also the photo, and the page face not detected in Harmy helps. If the points appear but are shifted, the most likely cause is angle or tilt.
What to tell the patient when the photo needs to be retaken?
No alarm and no blame. The photo is part of the process, and retaking it is a measure of care.
What mistakes to avoid?
- Generating the projection with a "so-so" photo to save time.
- Using a social media or WhatsApp photo (compressed and filtered).
- Accepting a smile because "it looks nicer."
- Adjusting the ml to compensate for a bad photo.
- Showing the patient a projection from a bad photo without warning them.
The best investment in projection quality is not in the app, it is in the camera, the light and the minute of preparation before you press the button.
Frequently asked questions
What gets in the way of AI projection the most?
Strong shadow, a turned or tilted face, a marked expression, beauty filters and a blurry photo. All of these go into the starting image and the AI does not correct them: it only alters what was marked.
How do I recognize a bad photo before generating?
Check that the face is facing forward, whole and sharp, with a neutral expression, even light and no filter. If the app does not detect the face or the points land out of place, the photo probably needs to be retaken.
Does the AI correct the lighting or expression?
No. Harmy's prompt forbids changing expression, lighting, background and hair, and forbids skin retouching. The projection changes only the marked areas, based on the ml chosen.
What to do when the projection looks strange?
First review the photo. Almost always the problem is in the shadow, angle, expression or filter. Then check the marking and the ml. Retaking the photo usually solves more than adjusting the intensity.
Harmy images are illustrative simulations generated by AI from the patient's photo to support the conversation between practitioner and patient. They are not a promise of results. Indication, technique, product and volume are solely the decision of the licensed practitioner. This content is for information only and does not replace a clinical assessment.