USE WHEN
Use this manual when
- Projects with scripts, storyboards or shot lists
- Teams evaluating both hosted and open models
- Productions ready to record rework, usable-shot rate and delivery cost
PRODUCTION MANUAL · Model routing
Model routing for serial production: assign a pipeline to each shot, not one “best model”
Build a project routing sheet around inputs, audio, shot control, revision and traceability across Seedance 2.0, Veo 3.1 Lite, Runway Gen-4.5 and Kling 3.
DO NOT USE AS
Do not use it as
- Choosing a single tool from vendor demos alone
- Cross-model rankings without fixed assets, versions and reviewers
01
VERIFIED FACTS
What the evidence currently establishes
- 01
Seedance 2.0 officially supports mixed text, image, audio and video inputs for reference generation, editing and extension; its launch notes also flag multi-subject consistency, text rendering and occasional audio distortion as areas for improvement.
- 02
The Veo 3.1 Lite model card documents natural-language or image inputs, high-resolution video with audio, and distribution through Gemini API, AI Studio, Vertex AI, Flow and Workspace.
- 03
Runway positions Gen-4.5 around motion, prompt adherence and visual consistency; these are vendor claims and still require fixed-shot project testing.
- 04
Kling 3's official guide can establish available controls, but does not create a comparable quality score against other models.
02
REQUIRED INPUTS
Prepare before starting
- 10–20 representative shots spanning dialogue, two-person interaction, fast action, transitions and inserts
- The same canonical character and environment references, shot intent and negative constraints
- Target aspect ratio, duration, native-audio requirement and subtitle-safe area
- Product, exact version, territory, plan and test date
- One review sheet covering narrative accuracy, drift, physics, audio, retries and waiting time
03
WORKFLOW
Execution workflow
- 01
Group shots by input and delivery risk rather than genre alone.
- 02
Assign a primary and fallback route to each group, with explicit switch criteria.
- 03
Lock assets, prompt, version, aspect ratio and duration for the first test.
- 04
Save inputs, outputs, seed or job ID, elapsed time and cost for every run.
- 05
Use the same reviewers to assess narrative accuracy, continuity, audio and revision difficulty.
- 06
Promote only passing shot classes into the route sheet; retest after model updates.
04
DELIVERABLES & QC
Deliverables and acceptance
Required deliverables
- Shot-to-model routing sheet
- Hashed fixed-test asset pack
- Generation and rework ledger
- Retest record after version changes
QC checklist
- Exact product version and test date recorded
- Same inputs and reviewers used
- Vendor demos separated from project tests
- Fallback route defined for failed shot classes
- Waiting time and manual rework captured alongside spend
EVIDENCE BOUNDARY
What this manual does not prove
This guide provides a routing method, not a cross-model leaderboard. Official pages establish public capabilities, access and disclosed limits; production conclusions require fixed samples, fixed versions and a shared review method.