How do you use an AI agent as a director of photography for your film?

Spin up a DOP agent inside the invideo agent, load it with a visual-language treatment (camera, lens, lighting, palette, composition, movement), then brief each shot in plain on-set language — "35mm, motivated rim light, slow push-in, hold on him until he lunges" — and let the agent translate that into the prompt, model choice, and reference stack for every frame.

Start by giving the DOP agent the framework it will shoot against. invideo is an agentic video tool with every current video and image model and upscaler available behind one agent, so you direct in language and the agent routes to the right model.

1. Load a visual-language document as the DOP agent's context. Write a structured treatment covering camera, lens and aspect ratio, lighting source, color palette (named tonal modes with hex values where you care about reproducibility), composition, movement, atmosphere, mood, and film/DP attribution — plus negative prompts of what to never do. One documented horror short ran on a 25-page director's treatment and the agent held it across the entire film; another encoded 14 sections of a director's grammar and a fixed 9-element prompt assembly order so every shot was built the same way. Upload once, tell it to save to context, and it stops drifting between shots. As Hridaye, invideo's creative director, puts it: "camera continuity carries from the treatment doc forward. you're not telling the agent how to move the camera every time. you set it once. it holds. that's the flow state."

2. Initialize a creative producer agent above it. Before the DOP agent shoots anything, a creative producer agent holds the full script, shot breakdown, and character details so every cinematography decision is grounded in story context. The DOP agent inherits that vision instead of guessing scene intent.

3. Brief shots like you brief a DP on set, not like you prompt a model. Conversational, specific, in cinematography vocabulary: lens, movement, motivation, blocking, where to hold and where to cut. "I want to stay on the feral guy when we run this scene. No back and forth cutting. We hold on him right up till he lunges" — that exact line produced the intended shot where manual prompting had failed. The agent compiles your direction into the structured parameters (focal length, movement vector, color temp, lighting plan, negative prompt) and the prompt itself — you stay in the director's chair.

4. Make the agent surface gaps instead of guessing. A real DP asks questions. Instruct yours to do the same: when you call a reverse, it should flag what's behind the actor if that wall doesn't exist yet, and offer narrative-loaded options before generating. One session unblocked a shot on the first attempt this way after manual prompting had stalled. Same posture when you state a lens or lighting claim — tell it to challenge you. In one production the agent had logged "anamorphic," was questioned, and corrected to spherical 35mm at 2.40:1 hard matte for that director's actual format, before the error propagated.

5. Run multiple DOP agents in parallel, one per scene. Different scenes want different eyes. Assign a DOP agent per scene rather than one agent for the whole film, and for complex sequences put two DOP agents on the same scene simultaneously. Documented productions have run 6–8 specialist agents in parallel across separate project pages — that pace of iteration is the actual unlock, not the automation.

6. Lock visual continuity by passing context forward, not by re-prompting. Lock character sheets and environment references before any video generation — that single step prevents most consistency problems downstream. For continuous coverage, feed the prior shot (or its end frame plus the character and location references) back to the DOP agent so the next generation inherits camera movement, lighting, and spatial logic. Inside invideo this routes naturally — Seedance 2.0 reference-to-video carries character and location context across clips, Kling handles multi-shot sequences natively, Veo holds physical realism, and the DOP agent picks per shot. You don't pick the model; you describe the shot.

7. Use it as a quality gate. Tell the DOP agent to check every generation against the treatment before returning it, and to flag deviations. In one production the agent caught that Stage A shadows were leaning blue-green instead of neutral gray, pulled the rule from the document, and offered a warmer pass without being asked. After the rough cut, send the assembly back and ask what's working and what isn't — one session caught an entity-reveal shot running at the wrong emotional stage register, which the human editor had missed.

This is how a working DOP agent behaves: persistent context up front, on-set language in the middle, agent-surfaced gaps and self-correction throughout, locked references carrying continuity from shot to shot. What works depends on the film — a 90-second horror piece with five emotional stages wants a different DOP agent setup than a one-take vampire sequence across cities.

Watch some of these to see what works for you:

camera continuity carries from the treatment doc forward. you're not telling the agent how to move the camera every time. you set it once. it holds. that's the flow state.

— Hridaye, invideo's creative director