How AI "Cinematic Scroll" Websites Actually Work (Two Different Approaches)
October 10, 2026 · 5 min read
Two approaches, same visual effect
Scroll through a cinematic AI-generated website and a camera seems to glide through a scene as you scroll. Under the hood, there are two completely different ways tools build this — and within the more popular approach, two different ways to pay for it.
Approach 1: AI video, sliced into frames
This is how Draftly and Parallux both work, though they differ on where the video comes from. The pipeline looks like:
1. An AI video model generates a short cinematic clip 2. The tool extracts frames from that video 3. As the visitor scrolls, JavaScript swaps between frames — scroll position maps directly to a point in the video's timeline
It's a neat hack. It *feels* 3D because the underlying video had camera movement baked in. There's no actual 3D scene in the browser and nothing to rotate or click into — but it's a lightweight, reliable way to ship a premium-looking cinematic page.
Where the two tools diverge is *what gets generated per site*:
- Draftly generates a brand-new video from scratch for every single site. That gives you a one-of-a-kind result, but it means every site costs Draftly real video-generation compute, which is why it charges before you ever see your output.
- Parallux matches your brand's prompt to a library of pre-generated cinematic scenes instead of generating fresh video per site. The AI model's job is picking the right existing scene (mood, palette, camera path) rather than rendering a new one. That makes generating and refining free, with the $25/mo cost only kicking in when you publish.
Approach 2: Real-time 3D rendering
This is how Vulk works. The pipeline looks like:
1. An AI model reads your prompt and outputs *parameters* — not a video, but a structured description: color palette, mood, which 3D objects or scene type to use, a camera path 2. A 3D engine (Three.js, running in the browser via WebGL) renders an actual 3D scene from those parameters 3. As the visitor scrolls, the *camera itself* moves through real 3D space in real time
This is more technically demanding to build, but the result is a real, interactive, lightweight 3D scene — not a pre-baked video. It doesn't require downloading a large video file, and it can in principle support real interactivity (rotating objects, clicking into details) that a frame-based video never can.
Why does this matter to you?
If you're picking a tool, it affects:
- Whether you can preview for free — a shared-scene-library approach (Parallux) can let you generate and iterate at zero marginal cost; a fresh-video-per-site approach (Draftly) generally can't
- How unique your output is — a brand-new generated video is one-of-a-kind; a matched scene from a library is shared with other sites that land on a similar match
- Interactivity — only real-time 3D (Vulk) supports actual rotation/interaction; both video-based approaches are scroll-only
Neither approach is "wrong." If you care about what's actually happening in the browser and whether you're paying for a fresh render or a shared template, now you know what to look for.