Best source photos for face swap: pose, light and resolution guide

The best source photo is a useful identity reference for the actual target, not necessarily your favorite selfie. This guide helps you select that input before you spend time changing settings or switching tools. The examples below are illustrative selection decisions; no swaps were generated or scored for this article. Current Magic Hour product pages and API documentation were checked on October 6, 2026.

Source versus target: label the files before uploading

Source means the face you want to insert. Target means the existing image or clip receiving that face. If you want your face in a permitted portrait scene, your selfie is the source and the portrait scene is the target. Reversing these files changes the job: the selfie becomes the scene and the other person becomes the replacement identity.

On Magic Hour Face Swap Photo

“Upload original photo” means the target. “Upload replacement face” means the source. For video, “Upload base video” means the target clip and “Upload face photo” means the replacement identity. Read those labels instead of assuming that the first upload slot is always the source.

For footage, start directly at Magic Hour Video Face Swap

If you are moving from a local face-swap workflow, keep the same permitted source and target while checking the browser version. Translate your file roles into the visible uploader labels. Do not change the source, target and production route simultaneously, because then a different result will not tell you which change mattered. For the complete upload-to-download sequence, use our Magic Hour video face-swap walkthrough

Inspect the face, not the photo’s overall size

Open each candidate at a size where you can inspect the eyes, nose, mouth and facial outline. Look for visible detail rather than a large file or impressive camera specification. A landscape-sized photo with a tiny person may contain less usable face detail than a modest portrait. Zooming into a face also makes compression blocks and motion blur easier to recognize.

Prefer an original or a good-quality export over a screenshot of a screenshot. If you have both, compare the face region before choosing. A file being accepted by an uploader proves format compatibility, not that it contains a readable identity reference. Likewise, an upscale is not evidence that missing identity detail has been recovered accurately.

Avoid judging a reference only from a thumbnail. Check whether eye shape, nose edges and mouth detail remain visible when enlarged. If your source is visibly smeared or pixelated, look for another original before attempting aggressive sharpening. The selection question is whether this person is clearly represented, not whether the image can be made to look crisper.

Choose pose, expression and light for the actual job

Magic Hour’s photo documentation recommends clearly visible features, compatible source and target lighting, and avoiding large differences in head angle. Its video guidance favors a front-facing or slightly angled source with a neutral expression. These are practical starting conditions, not a promise that every compatible pair works or a universal angle requirement across all tools.

For a still target, compare where the face points and how much of each cheek is visible. Start with a reference that does not force a dramatic pose change. For video, inspect the range of poses in the clip rather than choosing a reference that matches only its first frame. Do not select an extreme profile source solely because one target frame is a profile.

A neutral or moderate expression makes a useful starting candidate. A broad grin, tightly closed eyes or exaggerated mouth shape may be a poor identity reference for a calm portrait. However, the target still supplies the performance: choosing a smiling source is not a reliable way to make a neutral target smile or change a video’s spoken expression.

Check whether important features disappear into shadow or glare. Prefer a candidate where you can inspect both eyes and the mouth without a distracting lighting effect. Keep another suitable reference if the target has different light. Evaluate the eventual blend in the target scene; a pleasant-looking source portrait alone cannot establish that its lighting fits.

Keep an unobstructed reference and a sensible crop

Sunglasses, a mask, a hand or hair across the face can hide identity cues. If the person wears accessories in the target, that does not automatically mean the source should hide the same features. Begin with a reference where those features are visible, then inspect whether the actual target obstruction is preserved after processing.

Use a crop that makes the intended person unambiguous without cutting into the chin, forehead or cheeks. If another face is present, choose a single-person original when possible. Keep the uncropped file, then compare a moderate crop if detection fails. A crop cannot reveal a face that is covered, nor should it be treated as a guaranteed detection fix.

Do not assume the source’s hairstyle, ears, outfit or background will transfer with its face. A face swap is different from replacing the whole head or body. Keep enough facial outline to judge the reference, but evaluate success as a face-identity change inside the target scene rather than demanding that every surrounding source feature appears.

Build a small reference set without confusing Multi mode

Keep an approved primary portrait and a few genuinely different alternatives: perhaps another clear slight angle, expression or lighting condition. Give them descriptive names such as source-front-neutral and source-slight-angle-softlight. Extra copies of the same blurred selfie do not create useful alternatives. Select among references for each attempt; do not assume a single swap blends a folder of views.

Magic Hour’s single workflow accepts a replacement face image. Multi mode maps different target people to replacement faces; it is not a control for uploading several angles of one person into a single identity. In the APIs, all-faces uses one source for every detected target face, while individual-faces uses explicit mappings. Check the intended mode before a group-photo job.

Illustrative source-selection examples

These choices organize a first attempt; they are not measured quality rankings. A reference can pass the input inspection and still fail on a particular target. Keep the target, selected person and output requirements fixed when comparing the alternate source. Record why you changed it: visibility, expression, lighting or identity likeness.

Use different acceptance checks for photos and video

For a photo, inspect the complete exported image at the size where it will be used. Compare the intended identity, eye and mouth structure, skin transition, facial outline and any nearby glasses or hair. Also check that the right person changed and that surrounding target objects remain acceptable. Do not approve a crop that conceals a boundary problem in the full image.

For video, choose a short representative segment before a long render. Include the actual difficult moment, such as a head turn or hand crossing, if it matters to the final job. A clean frontal frame is a useful reference point, but it does not establish that a moving clip maintains identity through time.

Play the downloaded clip normally, then inspect the affected moments frame by frame. Watch for a face changing identity, disappearing, sliding or flickering during motion. Look at the neighboring frames before attributing the issue to the source. The original target may already contain blur, an obstruction or a cut that explains why that moment is difficult.

The source is a reference, not a repair instruction for every target frame. If the target face becomes tiny, hidden or heavily blurred, choosing another portrait may not solve the underlying visibility problem. Consider a different target segment, a tighter editorial shot or a manual edit rather than assuming the tool can infer everything behind an obstruction.

Troubleshoot the input layer before changing tools

The source face is not detected

Confirm that the replacement photo is in the replacement slot and contains the intended readable face. Check the actual returned message. Both Magic Hour API guides document no_source_face as a failure to detect a source face. Try a clearer single-person image before treating it as a general outage; a request-format or account error calls for a different fix.

The result is recognizable but looks wrong

Compare the source with the target’s pose and lighting, then inspect whether the source has obvious filtering or a very different expression. Change one candidate at a time. If the same defect remains with a clear alternate source at the same target moment, investigate the target or workflow capability instead of continually collecting more selfies.

Only one turn or obstruction fails

Compare that moment with the original video. Does the target face disappear behind a hand, become blurred or turn away? Retain a nearby clear frame as context. A different reference may be worth one controlled attempt, but do not present that as a guaranteed repair for unseen target features. If the shot is essential, plan finishing work or choose a better shot.

The wrong person changes

Check face selection and mapping before replacing the source. A perfectly clear reference still produces the wrong edit when it is assigned to the wrong target person. For a group, identify each intended original face and its replacement explicitly. Keep consent and file labels aligned so that similarly named portraits are not accidentally exchanged.

Prepare a source packet you can reuse

• Primary source: clear replacement identity, descriptive filename, original file preserved.

• Alternates: only references with a meaningful difference in pose, expression or lighting.

• Target: exact photo or clip segment, selected person and surrounding scene identified.

• Permission: allowed identity, input media, audience and intended distribution confirmed.

• Acceptance: recognizable identity, acceptable facial structure and blend, correct person, and video continuity when relevant.

Use media and likenesses you are permitted to transform. Keep that approval with the source packet rather than relying on a filename or the fact that a photo is public. Check the current product terms for intended commercial distribution. A subscription or disclosure does not independently settle permission to use someone else’s identity or footage.

For selecting a browser workflow after preparing your source, compare the face-swap photo tools. Keep the same source and target when evaluating a different tool so the input does not become another uncontrolled variable.

Official sources and product references

https://docs.magichour.ai/tools/image/face-swap-photo

https://docs.magichour.ai/tools/video/face-swap-video