Elliot Quartz

Back to work

Visual Cover Production Engine

A shot-for-shot AI filmmaking system that uses a human-directed source as a score, then lets Codex manage identity, generation, exact conform, and surgical repair.

August 2026 · Software / systems · AI Film, Codex, Generative Video

Cleopatra and Mark Antony walk together through the desert in warm backlight during an early shot of the visual cover film.

Response

  • Latest cut: v09
  • 37 shots / 1,068 frames
  • Native speed, no retiming

Details

Stack: Codex, GPT Image, Kling 3.0, FFmpeg

Generative video can make an impressive five seconds. It is much less reliable at sustaining a complete visual language.

The Visual Cover Production Engine solves that by treating an existing human-directed film as a score. The source provides the composition, blocking, light, camera movement, pacing, and cut points. I recast it with historical figures and direct the translation. Codex keeps those decisions coherent across image generation, motion, editing, and repair.

Cleopatra + Antony, Final Film v09

The complete 44.6-second Cleopatra VII and Mark Antony historical cover. This public edition is silent.

The film translates all 37 source shots into a late-Ptolemaic world around 34 BCE. Cars become chariots, a private jet becomes a Nile galley, and electric light becomes sun, fire, moon, or reflected water. The world changes while the source film's visual logic stays intact.

Original and Remake, Synchronized

Original Vogue film at left; Cleopatra + Antony at right. Both uncropped frames are synchronized across all 1,068 editorial frames.

The side-by-side is the clearest explanation of the method. The remake inherits authored rhythm and specificity without asking a model to invent taste from scratch. The source is the score; the historical world is the interpretation.

How the Pipeline Works

  1. Map the film. Every shot gets an exact timecode, frame count, composition, camera, light, action, and transition contract.
  2. Lock the translation. Identity, wardrobe, environments, and historical replacements are approved before dependent shots begin.
  3. Build privacy-safe guides. Pixelated source frames preserve pose and geometry without passing the original cast's facial details into GPT Image.
  4. Generate and animate. Start frames must pass likeness, anatomy, lighting, and composition checks before Kling receives a bounded motion brief.
  5. Conform and repair. FFmpeg rebuilds the original timing at native speed. Contact sheets and decode checks isolate weak shots so accepted work stays untouched.

The result is a recoverable production system, not a loose folder of prompts.

Surgical Revision

One three-shot sequence still looked like a modern luxury vehicle. I moved the same blocking into a Ptolemaic chariot house and regenerated only S18-S20. Those 67 frames changed; the other 1,001 editorial frames remained on their accepted lineage.

A labeled midpoint contact sheet showing all 37 shots in the final Cleopatra and Mark Antony historical film.
All 37 shot midpoints on one review surface. The repaired chariot-house sequence is S18-S20.

Human Direction, Agent Production

I directed the source, cast, historical translation, likeness standard, and final judgments. Codex carried those decisions through GPT Image, Kling, downloads, conform, recovery, and verification.

The finished cut contains 37 shots and 1,068 frames at 24000/1001 fps, with no retiming or interpolation. Human taste stays in charge while the agent preserves production memory and repetitive precision.

Pilot 00: Dalí + Picasso, Full Film

The first completed pilot recast a 44-second GQ fashion film with Salvador Dalí and Pablo Picasso. It established the same pattern: identity anchors, exact timing, retained versions, and shot-level repair.

The complete 44.6-second Dalí and Picasso pilot, presented as a silent visual cover.

Two completed covers show that the reusable work is the protocol: source-led direction, controlled identity, exact editorial structure, and evidence-based repair. The comparison footage is presented as a silent process study with source attribution.