Video Generation

Video at scale.
Routed right.

Four video models in production. Each handles a different class of output. The routing layer decides which model touches which brief — no single model handles everything.

Flagship — Seedance 2.0

Real output, not a placeholder

Veo — environment scene
Model breakdown

Four models. Four jobs. No overlap.

ByteDance (direct partnership)

Seedance 2.0

  • Multi-shot narrative sequencing
  • Dialogue + audio (SFX, ambient, BGM)
  • Cinematic camera language
  • Real-face reference blocked — character abstraction only

720p / 5s = 205 credits. Multi-shot syntax: [00:00-00:03] per segment.

Routing rule

Default first choice for narrative and cinematic content.

Seedance 2.0 — cinematic output frame

Kuaishou

Kling 3.0

  • Talking head + lip-sync (Omni mode)
  • Vehicle physics and fluid dynamics
  • Lifestyle b-roll and product loops
  • Motion control precision

Omni mode for lip-sync. Always end prompts with "then settles" for clean loop.

Routing rule

Talking head, lip-sync, product loops, physical simulation.

Kling 3.0 — bar scene, character detail

Google

Veo 3.1

  • Photorealistic environments
  • Complex scene composition
  • Three endpoints with different quality/cost tiers
  • Strong for documentary and editorial style

3 endpoints mapped. Best for photoreal scenes where character consistency is less critical.

Routing rule

Environmental scenes, documentary aesthetic, high-realism briefs.

Veo 3.1 — environment / scene output

xAI

Grok

  • Fashion and lifestyle editorial
  • Fast iteration on brand/product concepts
  • Natural outdoor and location shoots
  • Loose creative-brief interpretation

Best used for concept exploration before committing a brief to a slower, higher-fidelity model.

Routing rule

Fashion, lifestyle, and editorial concepts; early-stage creative exploration.

Grok — fashion lifestyle scene
Pipeline anatomy

From brief to CDN without touching it.

01

Brief intake

Shot description, character refs, tone, and output spec. Parsed by n8n into structured generation parameters.

02

Model selection

Routing layer assigns model based on task class — narrative, lip-sync, environment, or loop.

03

Prompt construction

Prompt DSL applied: shot language, timing brackets for Seedance, anchor mechanics for Kling.

04

Generation + retry

API call with retry logic and fallback routing if the primary model hits rate limits or quality threshold.

05

QC gate

Duration check, resolution verification, output scoring. Auto-reject and re-queue if thresholds not met.

06

CDN delivery

Cloudflare R2 upload, URL registered in Supabase, notification to downstream consumer.

Pipeline

Brief to CDN, unattended

Brief
Model select
Generate
QC
CDN
100%
Automated
4
Video models
QC gate

Output scoring

99.2%
Run success
Auto
Reject + re-queue