Generating Stacked Composite Images with Gemini AI

Stacked generation remains valid in multiple forms.

Approach A: Flexible Grid Strategy (General)

Useful when scene count varies widely and you want adaptive layouts.

• Compute cols, rows from scene count.
• Allow layouts like 2x2, 3x2, 3x3, 4x3.
• Crop output back into scene images with Sharp.

Approach B: Deterministic Pipeline Grid Strategy

Useful when downstream splitting and indexing must stay predictable.

• Use fixed 3x3 composite batches.
• Generate with gemini-3-pro-image-preview.
• Use deterministic dimension maps per videoType and imageSize.
• Persist grid_image and split slide_image assets with stable asset_index.

Empty-cell Handling

Both approaches should explicitly specify filler content for unused cells to keep crop positions stable.

Splitting Notes

• Sharp-based cropping is valid for standalone flows.
• Sandbox/ImageMagick splitting is valid for memory-constrained worker pipelines.
• When reading split binary files from sandbox, use byte mode to avoid PNG corruption.

Cost Perspective

Stacked composites are often cheaper operationally than one request per scene, especially with larger scene counts.