Make each identity easy to read
Use one adult in each photo. A clear, front-facing face gives the model more visible information than a tiny face, side profile or filtered screenshot. Keep the eyes, nose and mouth unobstructed.
Avoid sunglasses, masks, strong backlighting and group photos. Do not use a public figure’s portrait or a photograph you do not have permission to upload.
Leave room around the subject
Include the head, shoulders and enough of the outfit to show its shape. A full-body image can supply more clothing information, but only if the face remains sharp and sufficiently large.
Do not enlarge a very small image just to increase its pixel dimensions; it does not restore facial detail. Crop away other people while keeping comfortable space above the head.
Choose a compatible pair
The two photos can come from different locations. Similar lighting, a comparable camera angle and an uncluttered background make a more consistent starting point.
Distinct clothing can make the two performers easier to follow. Elaborate patterns, small writing and jewelry may change during motion. If an outfit must remain exact, AI dance generation may not be suitable for the job.
Check the files before uploading
Use JPEG, PNG or WebP, no larger than 8 MiB per photo. If a phone saves HEIC, export a JPEG first. Use an actual image file instead of renaming the extension.
Open both photos at a normal viewing size before selecting them. If you cannot see the eyes clearly, the model has little detail to work with.
Good photos help; they do not guarantee fidelity
Motion generation can change faces, fingers, clothing and body proportions. Two people crossing or overlapping can make identity consistency harder. Review the whole result before sharing.
This checklist is editorial guidance based on the input requirements and known limitations of image-guided video. It is not a claim that we have measured a success rate for each photo style.