Models and providers
Each uploaded photo is first turned into a colored-pencil character reference with OpenAI GPT Image 2 through fal. The video is then generated with ByteDance Seedance 2.5 through RunningHub, using the two references and a silent motion template. If the clip comes back one frame long, fal trims it. Our own server adds the soundtrack and checks the final file.
Current status
Generation is switched off on this site at the moment. No photos are sent to any AI provider until it is switched on.
How a video is made
Your two photos guide the appearance of the two people. A fixed scene template provides the elevator setting, framing, light and the sequence of the moment. The output is one 19-second 16:9 MP4. The template’s original soundtrack is added after generation; it is never sent to the AI model.
What is not used
No footage from (500) Days of Summer and no likeness of the film’s actors are part of the template or the output. People in the motion template are replaced by the two people in your photos.
Prompt safety check
Before any AI model is called, the generation prompts are checked with Waffo’s content-safety scan, which focuses on child-safety and other restricted content. Generation only continues when the scan allows it; a prompt that needs review waits, and a blocked prompt is never generated. Only the prompt text is sent to this check, never your photos, and we keep only a fingerprint (hash) of the prompt with the result. Photos and finished videos are still reviewed manually before delivery.
Known limitations
Faces, hair, hands, headphones and motion can change. Timing is approximate. An exact likeness, accurate speech and lip sync are not guaranteed.
Labelling
Please label shared videos as AI-generated. Do not present them as real footage of the people shown.