Lore

Blender gray previs as video input

A plain gray Blender animation, with no colour or materials and only geometry, motion and camera, is attached to Higgsfield as the video input. Words describe camera and timing slowly and imprecisely. The previs locks them.

"The video guides the movement and camera, while the images show how everything should look."

Workflow

  1. Animate the shot in Blender and adjust the camera alongside the animation.
  2. Use squash and stretch on the wheels, which removes the stiffness of rigid geometry.
  3. Export the gray preview.
  4. Attach it to the generation as the video input, together with image references for the taxi, trucks, characters, road and light.
  5. Write the prompt with the five-block-video-prompt-structure.
  6. Check the output with the generate-and-compare-review-loop.

Division of labour

Failure mode A textureless gray previs can be misread when the camera is too close: see zoom-out-when-the-model-misreads-a-reference.

Limit Previs fixes motion and camera but does not solve subtle acting: see Subtle character acting: the performance gap. Related: AI rendering trade-off: speed for control (previs use case), AI video inside a traditional animation pipeline.

A gray previs can be readable to its makers and unreadable to the model

The close Blender camera looked fine to the team because they knew the scene. Without textures or faces, though, the frame was mostly large gray shapes, so the model could not tell how characters and objects were positioned. Judge the previs as a model would see it, not as its makers do.

See previs-detail-two-sided-cost-for-performance for how adding detail also has a cost.