16. Optimization & cost#

Your CRM is now triggering renders automatically (chapter 15) — possibly hundreds per day. Each render consumes credits and takes wall-clock time. This final chapter covers four levers that bring down both: cache, scene granularity for parallel rendering, quality, and asset reuse.

Prerequisites: chapter 15. Optimization assumes you understand the rest of the pipeline.

Lever 1 — Caching#

Every renderable thing in JSON2Video is cached by default. The cache key is "everything that affects the output": the JSON of the element, its sources, its settings. If you submit a movie with the same JSON twice, the second render is served from cache — essentially free and instant.

To opt out for a specific element, set cache: false. Useful when:

• A source URL has the same path but changing contents (e.g. a price tag served by your CMS).
• You need a fresh AI generation to try different seeds.

{
  "type": "image",
  "prompt": "A craftsman house at sunset, magazine style",
  "model": "flux-schnell",
  "cache": false
}

In production, leave caching on. Each AI-generated image in the listing costs credits the first time — every subsequent render of the same listing is free.

Lever 2 — Split into more scenes#

JSON2Video renders scenes in parallel. The total render time is roughly max(scene durations), not the sum. Two effects:

• Splitting one 30-second scene into 6×5-second scenes is significantly faster.
• A monolithic scene with 30 elements is the worst case — it cannot parallelise.

In our listing, the chapter-13 iterate pattern already produces N scenes — one per room. That's optimal. If you have a long "stats roll" with many beats, break it into one scene per beat.

Lever 3 — Tune quality#

The movie-level quality field controls render fidelity:

low renders ~3× faster than high and costs fewer credits. A typical workflow: render low while drafting, then re-render high once approved.

{ "quality": "high" }

Lever 4 — Reuse AI assets across renders#

The most expensive operations are AI image and AI video generation. Two strategies:

4a. Hoist AI assets to preload. The preload array generates an asset once and exposes its URL as {{id_url}}, ready for every element that references it. The generation happens before the main render and is cached.

{
  "preload": [
    {
      "id": "twilight",
      "type": "image",
      "prompt": "A craftsman house at twilight, magazine real-estate photo, cinematic",
      "model": "flux-schnell"
    }
  ],
  "scenes": [
    {
      "elements": [
        { "type": "image", "src": "{{twilight_url}}" }
      ]
    }
  ]
}

4b. Upload the asset once via /v2/media. If you want full control, upload your AI-generated images to your own Media library and reference them by URL. No re-generation ever.

The complete final JSON#

{
  "resolution": "full-hd",
  "quality": "high",
  "cache": true,
  "client-data": {
    "listing_id": "L-4821"
  },
  "exports": [
    {
      "destinations": [
        { "type": "webhook", "endpoint": "https://your-app.example/json2video-callback" }
      ]
    }
  ],
  "preload": [
    {
      "id": "twilight",
      "type": "image",
      "prompt": "A craftsman house at twilight, magazine real-estate photo, cinematic",
      "model": "flux-schnell"
    }
  ],
  "variables": {
    "address": "123 Oak Street",
    "rooms": [
      { "name": "Exterior",       "image": "https://cdn.json2video.com/assets/images/sample-house-front.jpg" },
      { "name": "Chef's Kitchen", "image": "https://cdn.json2video.com/assets/images/sample-house-kitchen.jpg" },
      { "name": "Master Bedroom", "image": "https://cdn.json2video.com/assets/images/sample-house-bedroom.jpg" }
    ]
  },
  "elements": [
    {
      "type": "audio",
      "src": "https://cdn.json2video.com/assets/audios/uplifting-corporate.mp3",
      "volume": 0.4
    },
    {
      "type": "voice",
      "text": "Welcome to {{ address }}.",
      "voice": "en-US-EmmaMultilingualNeural",
      "start": 1.5
    },
    {
      "type": "subtitles",
      "language": "en",
      "settings": {
        "style": "boxed-word",
        "font-family": "Inter",
        "font-size": 90,
        "font-color": "#FFFFFF",
        "position": "bottom-center",
        "all-caps": true,
        "box-color": "#0E7C66"
      }
    }
  ],
  "scenes": [
    {
      "duration": 4,
      "elements": [
        {
          "type": "component",
          "component": "basic/000",
          "settings": { "headline": "FOR SALE", "subline": "{{address}}" }
        }
      ]
    },
    {
      "duration": 4,
      "transition": { "style": "fade", "duration": 0.5 },
      "iterate": "rooms",
      "iterate-as": "room",
      "elements": [
        { "type": "image", "src": "{{ room.image }}" },
        { "type": "text", "text": "{{ room.name }}", "position": "top-left", "x": 60, "y": 60 }
      ]
    },
    {
      "duration": 4,
      "transition": { "style": "fade", "duration": 0.5 },
      "elements": [
        { "type": "image", "src": "{{twilight_url}}" },
        { "type": "text", "text": "At Twilight", "position": "top-left", "x": 60, "y": 60 }
      ]
    }
  ]
}

Expected output#

Functionally the same listing as chapter 15 — but renders faster on first hit (parallel scenes), free on repeat hits (cache), with the AI asset hoisted into preload so it generates exactly once and is reused everywhere. Sample render: tutorial-16.mp4 (placeholder).

What you learned#

• cache: true (default) makes repeat renders of the same JSON essentially free.
• Splitting work into more scenes lets the renderer parallelise — total time tracks the slowest scene, not the sum.
• quality is a three-step dial; use low for drafts and high for delivery.
• Hoist expensive AI assets into preload so they are generated once and reused across every element.

You finished the tutorial#

You now have a production-grade real-estate listing pipeline: data-driven, conditional, narrated, captioned, cached, parallel, and webhook-delivered. Where to go next:

• Reference — every field and every endpoint.
• Guides — task-oriented walkthroughs (dashboards, no-code, advanced patterns).
• For AI Agents — set up an MCP server or hand these docs to your coding agent.

Previous chapter#

← 15. Webhooks