A face swap can fail even when the file opens normally and the person is easy for a human to recognize. The model still has to locate a face, estimate its pose, preserve expression and lighting, and blend a different identity into the original scene. A weak input makes every one of those steps harder.
This guide gives you a repeatable 100-point photo quality score. It is not a promise that a high-scoring photo will be perfect. It is a practical way to compare candidate images before you upload one and spend credits.
The 100-Point Scorecard
Score the target photo in five categories. Each category is worth 20 points.
| Category | 16–20 points | 10–15 points | 0–9 points |
|---|---|---|---|
| Face size and detail | Face is large and sharp | Face is usable but small or slightly soft | Face is tiny, pixelated, or heavily compressed |
| Pose and visibility | Front or mild three-quarter view | Moderate turn or tilt | Profile, extreme tilt, or face partly outside frame |
| Lighting and contrast | Even light with visible features | Some shadow or mixed lighting | Blown highlights, deep shadow, colored stage light |
| Occlusion and composition | Eyes, nose, mouth, jaw are clear | Minor overlap from hair, glasses, or hand | Face is covered, cropped, or overlapped by another person |
| Reference compatibility | Target and reference have similar angle and framing | Differences are noticeable but manageable | Angles, expressions, or facial scale are radically different |
Add the five scores:
- 85–100: strong first-choice input
- 70–84: likely usable, but inspect the weak category
- 50–69: test only if the image is important and cannot be replaced
- Below 50: find or create a better source image first
The score is most useful when you compare two or three photos of the same subject. It turns “this one looks clearer” into a decision you can explain.
1. Face Size and Detail: 20 Points
Open the image at 100% zoom. Do not judge it only as a thumbnail.
Give 18–20 points when the face occupies a meaningful part of the frame and individual features remain sharp. You should be able to see the boundaries of the eyes, lips, nostrils, and jaw without guessing.
Give 12–17 points when the face is clear but moderately small, or when the image has mild softness. A modern phone portrait taken from several feet away often lands here.
Give fewer than 10 points when the face comes from a distant group photo, a screenshot of a social post, a video frame with motion blur, or an image that has been repeatedly downloaded and recompressed.
A practical check is to crop around the head and view that crop by itself. If the crop already looks blocky before processing, the model cannot reconstruct real detail that is not present. Upscaling may make the file larger, but it does not restore authentic eyelashes, skin texture, or edge information.
2. Pose and Visibility: 20 Points
Face-swap models work best when they can estimate where the major features sit in three-dimensional space.
A front-facing portrait is not the only usable input. A mild three-quarter angle can look more natural because it preserves depth. The problem begins when the far eye nearly disappears, the nose crosses the cheek outline, or the jaw is hidden by the neck or shoulder.
Use this scoring guide:
- 18–20: front-facing or a mild three-quarter view; both eyes visible
- 14–17: noticeable head turn, tilt, or upward/downward camera angle
- 8–13: strong turn with one eye or part of the mouth difficult to see
- 0–7: near-profile face, extreme tilt, or only part of the face in frame
Expression also matters. A relaxed or moderate smile is usually easier than a wide open mouth, tightly closed eyes, inflated cheeks, or a face captured mid-speech. Those expressions are not forbidden; they simply require more precise geometry.
3. Lighting and Contrast: 20 Points
The model must reproduce the lighting of the target scene. If the target has soft window light but the reference has harsh flash, blending becomes harder. The same is true when half of the target face is nearly black and the other half is overexposed.
High-scoring lighting has three qualities:
- Facial landmarks remain visible.
- Skin is not clipped to pure white or crushed into featureless shadow.
- The color cast is consistent across the face.
Mixed light deserves special attention. A face may be lit blue from one side and orange from the other at a concert or event. A human interprets the scene correctly, but the swap may produce a pasted-on patch with neutral skin color.
Do not automatically brighten the entire image. First inspect the histogram or use a simple exposure control. Recovering moderate shadow can help; aggressively lifting a nearly black face usually exposes noise and compression blocks.
4. Occlusion and Composition: 20 Points
Occlusion means that another object covers part of the face. Common examples include:
- sunglasses or reflective lenses;
- hair across an eye or cheek;
- a microphone over the mouth;
- a hand touching the chin;
- a hat casting a hard shadow;
- another person overlapping the jaw; and
- text, stickers, or watermarks placed across facial features.
Thin eyeglass frames may work. Large mirrored sunglasses remove eye information and should score much lower. A small strand of hair is different from a curtain of hair covering half the face.
Composition also includes crop boundaries. A photo that cuts through the forehead or chin can force the model to invent an edge. Leave some space around the head when possible. You can crop tighter later, after reviewing the result.
If a group image contains several faces, the most reliable way to control the intended subject is to create a copy and crop around one person before uploading. faceswap does not provide a precise “face number” selector in the current workflow, so the detection result may not match your assumption. Our group-photo workflow explains this in detail.
5. Reference Compatibility: 20 Points
A face swap combines a target photo with a face reference. On faceswap, you can use the built-in reference or upload a permitted custom reference.
Compatibility does not mean the two people need identical head shapes. It means the model receives comparable visual evidence.
Compare these properties:
- head angle;
- camera height;
- facial expression;
- visible teeth or closed mouth;
- eyewear;
- apparent age and facial hair; and
- face scale within the image.
If the target looks sharply left while the reference faces straight ahead, the reference provides little information about the needed side view. If the target has a wide smile but the reference has a closed mouth, teeth and lip boundaries become more difficult.
When using a custom reference, choose a clear, close, front-facing portrait first. The reference should help the model understand identity. The target should help it understand scene, pose, expression, and lighting.
Three Worked Examples
Example A: Window-light portrait
- Face size and detail: 19
- Pose and visibility: 18
- Lighting and contrast: 19
- Occlusion and composition: 18
- Reference compatibility: 17
- Total: 91
This is a strong first attempt. No preprocessing is needed beyond confirming the crop and upload format.
Example B: Event photo from a phone
- Face size and detail: 14
- Pose and visibility: 15
- Lighting and contrast: 11
- Occlusion and composition: 13
- Reference compatibility: 14
- Total: 67
The image may work, but colored light and partial overlap create risk. Try a tighter crop and a reference with a similar angle. Keep the original file so you can compare rather than repeatedly editing the same export.
Example C: Screenshot of a distant group photo
- Face size and detail: 6
- Pose and visibility: 10
- Lighting and contrast: 12
- Occlusion and composition: 7
- Reference compatibility: 12
- Total: 47
Do not spend the first attempt on this screenshot. Look for the original photo, crop from the highest-resolution version, or choose a different frame.
A Three-Minute Pre-Upload Test
Use this sequence before opening the generator:
- View the original at 100% zoom.
- Crop a temporary copy around the intended head and shoulders.
- Confirm that both eyes, nose, mouth, and jaw are interpretable.
- Check for motion blur by looking at eyelashes, glasses, or hair edges.
- Compare the target angle with the built-in or custom face reference.
- Score all five categories.
- If the score is below 70, fix the lowest category or choose another image.
Do not run five random attempts when one deliberate input change would teach you more. Change one variable at a time: crop, reference, exposure, or source image. This makes it possible to identify what improved the result.
What Editing Helps Before Upload?
Useful preparation is conservative:
- straighten an accidental tilt;
- make a clean crop that isolates the intended person;
- apply mild exposure or white-balance correction;
- export once at a sensible quality; and
- remove an unnecessary border added by a screenshot tool.
Avoid aggressive beauty filters, oversharpening, artificial blur removal, face reshaping, or repeated JPEG exports. These can create edges that look sharp to you but are structurally false to the model.
For format choices, see JPEG vs PNG vs WebP for face swap inputs.
The Final Decision Rule
Use the highest-scoring photo that also has the right expression and context for the image you want to create. Technical sharpness alone is not enough. A perfectly sharp profile may still be worse than a slightly softer front-facing portrait.
After generation, inspect the output at both normal viewing size and 100% zoom. Check the eyes, hairline, jaw, skin transition, and any occluded area. If something looks wrong, use our regional artifact troubleshooting guide instead of submitting the same inputs repeatedly.
You can test a prepared image in the faceswap tool. Only upload photos you are allowed to process, and review our Content Policy before sharing a result involving another identifiable person.
