Implement computer vision solutions 3.1 Design and implement image- and video-generation solutions

A fitness brand uses Sora 2 (preview) to turn a still photo of its trainer into a short…

Study guide skill: Implement a solution that generates videos from text prompts and reference media

Question B7-001

A fitness brand uses Sora 2 (preview) to turn a still photo of its trainer into a short clip, passing the photo as input_reference. The photo is a JPEG whose resolution matches the requested output size. Every job with this photo fails, while the same request with a photo of the empty gym works. What's the most likely cause?

Select one answer.

  • A The photo must be passed as remix_video_id instead of input_reference when it shows a person
  • B The reference image must be smaller than the output resolution so the model has room to add motion
  • C The Sora 2 API currently rejects input images that contain human faces, so the reference image must not show the trainer's face ✓ Correct answer
  • D input_reference accepts only PNG files, so the JPEG must be converted first

Answer: C

The Sora 2 API enforces content restrictions: input images with human faces are currently rejected, and real people can't be generated, so a photo of the trainer fails while the empty gym works. JPEG is an accepted input_reference type, and the photo already matches the output size as the page requires. remix_video_id takes the ID of a previously completed video, not an image.

Deep dive

How it works

Sora 2 (preview) applies content restrictions to everything you send: prompts, reference images and transcripts. The rules on the page:

  • Only content suitable for audiences under 18 is generated.
  • Copyrighted characters and copyrighted music are rejected.
  • Real people, including public figures, can't be generated.
  • Input images with human faces are currently rejected.

That explains the pattern. The empty-gym photo passes. The trainer photo, same format and size, fails every time, because the reference image contains a face. The page's advice is to make sure prompts, reference images and transcripts respect these rules to avoid failed generations.

Practical options: use a reference image without a visible face, for example the trainer from behind or a product or location shot, and describe the person in the prompt as a generic character instead of a specific real person.

The rest of the request is already valid. input_reference accepts JPEG, PNG or WebP, and the image must exactly match the requested output size, which it does.

The other options, and when you'd use them

  • Pass it as remix_video_id: remix_video_id takes the ID of a previously completed video. It isn't for images.
  • Image smaller than the output: The rule is the opposite. input_reference must match the output size exactly.
  • PNG only: Accepted types are image/jpeg, image/png and image/webp. JPEG is fine.

Exam tip

Sora 2 job fails only when a person is in the reference image = human faces in input images are currently rejected. Format (JPEG/PNG/WebP) and exact size are the other two checks.

References

Last reviewed:

Ready to test yourself? The quiz asks this and every other AI-103 question, with instant feedback. Practice this in a quiz →

All Computer vision questions · AI-103 study guide

Thanks for your feedback!