Dailies

Inspiration

Previsualization — seeing a scene before you shoot it — is one of the most valuable steps in filmmaking and one of the most expensive. Studios spend weeks and tens of thousands of dollars on previs artists and studios so a director can make decisions before the costly shoot day. Everyone else shoots blind.

As a screenwriter, I wanted to see my scenes — cast, blocked, moving — without that cost. But every AI video tool races toward the same thing: "paste a script, get a finished film." That's the wrong tool for a filmmaker. A director doesn't want a black box that spits out a movie; they want a crew they can direct, and control at every step.

So I built Dailies: an AI film crew you direct.

What it does

A director hands Dailies a screenplay scene. A crew of specialized agents turns it into a watchable previz cut, with a human approving at every gate:

Cinematographer breaks the scene into a shot list — sizes, angles, camera moves.

Specification Gate (human-in-the-loop): the system pre-fills each character's identity, wardrobe, and props plus the location from the script, flags what it had to infer, and the director confirms or corrects before anything generates — so identity and setting don't drift.

Casting + Art Department generate a locked reference for each character and each set, then render stills conditioned on those references.

Camera animates the key shots into motion with lip-synced dialogue.

Story Supervisor checks continuity — wardrobe, location, appearance drift — before any money is spent on motion.

Editor assembles the approved shots, in the director's order, into one cut.

Video Reviewer watches the assembled cut, flags what's off, and tells the director exactly what to reshoot — closing the loop.

How I built it

Every agent reads and writes one shared state object, so the scene stays coherent across the pipeline. The Cinematographer and Story Supervisor are built natively on Google's Agent Development Kit; the full stack runs on Vertex AI with Gemini 2.5 Flash, Gemini 2.5 Flash Image, and Veo 3. Character consistency is conditioned on a generated reference image per character and measured with a face-embedding metric (InsightFace), not asserted. The whole thing was built in IBM Bob, which caught and fixed its own bugs along the way.

Challenges I ran into

The hardest problem was character consistency — the thing that makes AI-generated scenes look broken, because the same character's face changes from shot to shot. I solved it by having the Casting Director generate one canonical reference image per character, then passing that image as actual multimodal input to every shot the character appears in — not just a text description. The result: the same recognizable person across every shot. [Include the reference + two-shot comparison here.]

What I learned

The right unit isn't a perfect first-pass render — it's a fast loop with a critic in it. The Video Reviewer flagging a continuity issue and the director reshooting it is what makes Dailies a workflow instead of a novelty.

What's next

Feature-length consistency, PDF screenplay upload, the full crew (dialogue, score, editorial assembly), and putting it in front of working filmmakers — because the real value isn't "AI makes video," it's collapsing the cost of previs for the people who currently pay a fortune for it.

Built With

• agent-engine
• gemini
• google-adk
• google-cloud
• ibm-bob
• imagen
• python
• veo
• vertex-ai