How Do You Know if AI Video Followed Your Script? This Tool Shows You.

Taruma Sakti Megariansyah built SceneFlow, a free, open-source tool that syncs a screenplay or prompt script against AI-generated video on a shared timeline, so you can see exactly where the model followed your instructions and where it drifted. Version 2.3.0 shipped this week with a genuinely useful set of workflow upgrades.

The Actual Problem It Solves

Evaluating whether an AI video model followed a detailed prompt usually means watching the clip once, rereading the prompt, and trying to hold both in your head at the same time. SceneFlow replaces that with a synced, color-coded timeline: your script scrolls alongside the video, highlighting which line is "playing" at any given moment across eight cue categories, dialogue, action, camera, shot, audio, VFX, transitions, and environment.

Cue creation isn't automated. You either manually highlight text and set timestamps yourself in Edit Mode, or generate the sync data externally using a multimodal model like Gemini to analyze the footage against your script's timecodes first.

What's New in 2.3.0

The multi-track sync timeline is the headline addition: horizontal lanes with a continuous timecode ruler, a stationary playhead, and zoom presets for working at different levels of detail. A docked inspector panel adds one-click replay on individual cues alongside an 8-slot category indicator strip for quickly scanning what's active.

A few smaller but genuinely thoughtful additions round out the release: a pure black canvas mode for clean screen-recording compositing in Premiere or Resolve, a colorblind-safe mode addressing protanopia and deuteranopia specifically, and a draggable workspace with a calibrated split ratio between script and video panes.

A Second, More Technical Scripting Format

Beyond standard screenplay formatting, SceneFlow supports something the developer calls "Auteur Script," a structured, state-based prompting format specifically for reasoning-heavy video models.

Instructions get organized into modular blocks, Intent, Logic, Aesthetic, Opening, and Execution, with each shot treated as a "Macro-State" that explicitly inherits spatial and physical continuity from the state before it, aimed at preventing a model from silently resetting geometry or lighting between generated segments.

That's a genuinely different approach to prompt engineering than a single paragraph of natural language, closer to a technical spec than a screenplay, built specifically for the continuity problems that plague longer AI-generated sequences.

Competitive Context

This is a small project, 13 stars on GitHub as of this writing, built by a solo developer rather than a company. That scale matters for setting expectations: this isn't a polished commercial product, and cue creation still requires real manual effort or a separate AI analysis step.

But it's solving a genuine, specific gap. Most AI video discourse focuses on generation quality; SceneFlow is aimed squarely at the evaluation side, actually measuring whether a model did what you asked, which is a less glamorous but genuinely necessary part of serious AI filmmaking work.

The Signal in the Noise

Tools like this tend to matter more than their star count suggests, because they solve a workflow problem serious practitioners run into repeatedly rather than chasing a flashy demo.

If you're testing the same script across multiple video models, or trying to systematically identify where a specific model tends to drop instructions, having a structured way to compare prompt against output is genuinely useful groundwork, even if the tool itself stays niche.

Would a structured prompt-adherence workflow like this change how you evaluate AI video output, or does eyeballing it against your script tend to be good enough for your own work?

The Details

• Tool: SceneFlow, open source (MIT license), built by Taruma Sakti Megariansyah
• Core function: syncs a script or screenplay against video on a shared timeline, highlighting matching cues in real time
• Cue types: dialogue, action, camera, shot, audio, VFX, transition, environment (8 total)
• New in v2.3.0: multi-track sync timeline, docked cue inspector, collapsible video track, pure black canvas mode, colorblind-safe mode, draggable workspace
• Cue creation: manual (highlight and timestamp) or external, via multimodal AI analysis of footage against script
• Scripting formats supported: traditional screenplay format and "Auteur Script," a structured, state-based format for reasoning-heavy video models
• Access: free, live web app, source available on GitHub

Resources & Reads

• SceneFlow live app
• SceneFlow on GitHub
• SceneFlow's example catalogue