Video Generation and Post-Production Assembly Workflow

Learning objective

Treat a generation model as a supplier of shot material, then design the states, files, and failure boundaries from job submission through the editing handoff.

An auditable production pipeline

1. Lock the brief: record purpose, audience, aspect ratio, duration, rights, safety, and budget, then freeze the source_revision.
2. Preflight authorization: object-level authorization/ACL, intended use, retention period, and consent all pass first. Material that fails does not enter preview, scoring, or model calls.
3. Storyboard rehearsal: confirm pacing with text boxes or lawful still frames before expensive generation.
4. Submit one shot at a time: each shot has a stable ID, request hash, reference-asset inventory, and transform_id.
5. Manage asynchronously: retain the vendor job ID; update status by polling or webhook.
6. Ingest and review material: inspect MIME type, duration, dimensions, frame rate, audio tracks, hash, and provenance metadata. Only accepted material enters the approved bin; retry only the failed shot.
7. Post-production: edit, transition, color-grade, caption, voice, add sound effects, mix, and lay out graphics; append provenance at every step.
8. Review and release the finished piece: assess audio/video synchronization, continuity, rights, safety, Content Credentials, and release specifications. Assign a release_id only after human approval; factual external claims also need evidence_supported.

Video APIs are normally long-running jobs. For example, the OpenAI official page accessed on 2026-07-22 describes creating a job, checking its status or receiving completion through a webhook, then downloading its content. That page announced that the Sora 2 Videos API was scheduled to shut down on 2026-09-24; this is a future plan, not an event that has already happened. The business layer therefore needs its own state model and adapter, rather than treating one vendor lifecycle as a permanent contract.

flowchart LR
    A["brief + source_revision"] --> B{"object-level authorization/ACL"}
    B -->|"denied"| X["do not read, score, or submit"]
    B -->|"allowed"| C["shot list and adapter"]
    C --> D["asynchronous job"]
    D --> E["shot material + transform_id"]
    E --> F{"human quality/safety/rights review"}
    F -->|"rejected"| R["shot-level recovery or stop"]
    F -->|"approved"| G["post-production and finished piece"]
    G --> H["release_id, disclosure, release"]
    H --> I{"external factual claim"}
    I -->|"yes"| J["independent evidence meets evidence_supported"]
    I -->|"no"| K["release synthetic content"]
    J --> K
    B -. "revocation/deletion propagation" .-> E
    B -. "revocation/deletion propagation" .-> H

File and version rules

A useful filename pattern is project-shot-version-status.ext, for example demo-S03-v02-approved.mp4. Also retain a machine-readable manifest: input asset_id / source_revision / hashes, request hash, vendor ID, generation time, technical-probe results, transform_id, selection rationale, post-production chain, and release_id. Separate raw candidates, proxy files, approved material, and finished pieces into distinct directories; do not use “final-final2.”

Before transcoding or assembly, normalize canvas, frame-rate strategy, pixel format, audio sample rate, and time base. Concatenating files with different technical specifications directly can cause audio/video drift or player-compatibility failures. FFmpeg is a common tool, but commands must be generated from actual probe results; this course does not run commands against media that do not exist.

Idempotency and failure recovery

After a client timeout, query the retained job ID before creating another job. Download into a temporary file, validate it, then rename atomically. Each shot has a maximum retry count and a stop label; ACL denials, policy denials, and missing rights must not be retried automatically. Preserve the last approved version so a failed new edit or extension can roll back the material choice without destroying the original. On revocation or deletion, propagate invalidation through asset_id → transform_id → release_id to shot candidates, proxy files, finished derivatives, caches, and links, while retaining only the minimum audit record legally required.

Common mistakes and troubleshooting

• Saving only the final MP4: the request, inputs, and approval evidence are lost, preventing meaningful review.
• Duplicate webhook events: event processing must be idempotent and deduplicated by event/job state.
• Mixing generation and post-production responsibilities: precise captions, logos, and legal copy usually belong in post-production.
• Remaking everything after a full-piece failure: recover from shot-level checkpoints.
• Breaking provenance after transcoding: record the derivation relationship and output hash in the manifest.

Exercise and self-check

Draw the state diagram planned → submitted → running → generated → reviewed → approved/rejected → edited → published. Write a recovery action for each of: a duplicate webhook, interrupted download, incompatible technical specification, and human rejection.

Next: Audio, Captions, and Accessibility Interfaces.

References

• OpenAI Video generation guide (asynchronous workflow and planned future shutdown; checked/accessed 2026-07-22)
• FFmpeg Documentation (checked 2026-07-22)