AUTEUR: the first multi-model film

AUTEUR: the first multi-model film festival.

Demo: AUTEUR - The first multi-model film festival | Melius Challenge

Something I wish existed: a film festival where AI models compete as directors.

One sentence ("a child loses a balloon") given to six imaginary directors. Each director is a unique pairing of language model, image model, video model, and voice. Each interprets the same story in their own visual language.

THE PAINTER: Ghibli x Tarkovsky. Amber. Patient. (Flux Pro Ultra / Kling 3 Pro)

THE PROVOCATEUR: Lynch x A24. A backyard at the wrong hour. (GPT Image 2 / Kling 2.5 Turbo)

THE ROMANTIC: Wong Kar-wai. Step-printed rose. (Nano Banana Pro / Seedance 2.0)

THE DOCUMENTARIAN: Attenborough through tall grass. (GPT Image 2 / Wan 2.7)

THE MAXIMALIST: Luhrmann on a scarlet rooftop. (Flux 2 Max / Kling 2.6 Pro)

THE MINIMALIST: Bresson. No music. No voice. (Flux Schnell / LTX 2.3)

How I built it:

Drove Melius from an external agent via the MCP server.

Scaffolded the entire festival in a single agent call. 90 nodes, 85 edges, six labeled director groups, in-character manifestos and critic blurbs pre-filled, models pinned per director, edges auto-wired.

Locked visual DNA in-canvas with Mel. Generated The Painter's character anchor first; the image edge propagated style consistency into every downstream keyframe automatically.

Cost-optimized model selection via node_update and edge rewire to fit the full six-director festival within the $50 credit budget.

Directed final shots in-canvas with Mel for refinement.

Why only Melius:

Six different AI model families orchestrated on one shared canvas. Structurally impossible on Sora, Flora, ComfyUI, or Runway. The MCP let an external agent compose the entire 90-node scaffold in one shot; Mel directed the refinement. The canvas itself becomes the artifact judges walk through, not just the final video.

Canvas: https://app.melius.com/projects/bee36d9f-73b7-478f-8ca0-4d9e91f4adbd/canvas/eca1500c-b72e-4f2c-93fc-bfddeb4f0fd3

Feedback on Melius:

The MCP is the genuine differentiator. Scaffolding 90 nodes and 85 edges across six model families in a single agent call is something no other generative platform supports. Auto-edge switching (an image edge flips the target into image-to-image mode without manual config) is elegant. Mel and MCP create a two-handed creative loop: external agent for structure, Mel for taste. Small frictions: the auto-layout placed some director portrait nodes far from their group center; switching model families across the same node sometimes required manual edge rewiring because dstHandle names differ between variants. Overall: a real leap in agent-driven creative tooling. AUTEUR could not have been built this way on any other platform.

Shared on X: https://x.com/Jonatha98833942/status/2056111585236545569?s=20