Skip to content

StoryCore: multi-subject video generation router #40

Description

@zedarvates

Add a routing layer for multi-subject image-to-video/video generation so StoryCore can choose a generation strategy instead of forcing every scene through a single pass.

Strategies

  • DIRECT: generate the whole scene at once. Best for simple compositions and low interaction complexity.
  • PARALLEL: generate subjects independently, then composite/reconcile. Best when subjects are mostly independent and identity preservation matters more than cross-subject interaction.
  • SEQUENTIAL: establish scene/background, then add subjects/interactions progressively with validation between stages. Best for strong interactions or spatial dependencies.

Router inputs

Use shot specification plus lightweight analysis:

  • subject count
  • interaction graph / contact requirements
  • occlusion level
  • identity criticality
  • camera motion
  • temporal dependency on previous shot
  • available compute budget

Validation

Run existing/planned preview and temporal validators between sequential stages and after parallel composition. Bound retries and retain evidence/provenance for every branch.

Metrics

On a small fixed benchmark of multi-character shots compare:

  • identity drift
  • missing/extra subjects
  • interaction correctness
  • temporal/geometric drift
  • render time/cost
  • retry count
  • human preference/acceptance

Exit criteria

At least three multi-subject fixtures are routed automatically; for each, compare router-selected strategy against DIRECT baseline and record a reproducible run manifest.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions