A proof of concept in AI video production. We took low quality phone footage of a yoga session and rebuilt it as an HD, two-camera studio video: the model recreated from a five-panel character sheet, a 3D studio with the camera positions mapped out, and professional lighting, made with a mix of Claude and Higgsfield (Seedance 2.5). There are a few misalignments, as you'd expect at this stage, but with refinement it could be a great final product.
How we made it
Creating a character sheet of the model
- We created a five-panel character sheet of the real-life model to power the AI generation.
- It kept the person in the final video as similar to the real person as possible.
Designing a virtual studio
- We created the 3D space, with the camera positions mapped out.
- We used location and background images created for the shoot to show the lighting and scenery we wanted.
Planning a multicam shoot
- We set up a multicam shoot with two cameras in the virtual studio.
- We wrote a shot list with timestamps for when to cut to each camera.
AI video production in HD
- We ran her movements through the AI model, using a mix of Claude and Higgsfield (Seedance 2.5), to rebuild the video in HD.
- We changed her clothes, lighting and setting without a reshoot.
Results
HD, multicam, no reshoot
Low quality phone footage became a professional, two-camera studio video.
A proof of concept
A few moments don't line up yet. With refinement, we could get to a great final product.
