FilmFlow turns a text prompt into a Blender camera path

FilmFlow, a research system from Shanghai Jiao Tong University, Cornell University, UCLA and Shanghai Artificial Intelligence Laboratory, directs a camera through an existing Blender scene by writing an explicit trajectory and then looking at the render. The paper behind it, FilmFlow: Agentic System for 3D Virtual Film Cinematography, appeared as arXiv:2610.30753v1 on 22 October 2026.

Anyone who has chased a camera move through a video model knows the outline of the problem. A prompt can ask for a slow push-in, and the model can drift, overshoot or lose the subject halfway through. By the time the drift is visible, the render is paid for. FilmFlow's answer is to compute the camera move first, as numbers in a 3D scene, render it in Blender, inspect the result, and rewrite the numbers before any of it goes near a video model. A video-review write-up of the release frames the same pitch: models are promptable but not directable.

The center of the system is a Director agent that assembles each shot from 22 atomic camera operations. The library is a named list of film moves, grouped by what each one controls: position (Dolly In, Dolly Out, Truck, Pedestal Up, Pedestal Down, Orbit, Path Follow, Position Follow, Height Follow, Distance Follow), orientation (Pan, Tilt, Aim Follow, Screen Position Lock, Roll, Horizon Lock), framing (Zoom, Frame Subject), focus (Focus Pull, Focus Follow), plus Shake / Handheld and Static Shot. Each operation carries its own parameters and its own active window of time, and those windows are allowed to overlap. Orbit, Dolly In and Pedestal Up running together give a camera that circles a subject while approaching and rising, and each of the three keeps its own adjustable numbers for the Critic to argue with later.

Three agents support the Director. An Interpreter grounds the instruction in the scene's object list. A Scout hunts for the opening viewpoint before planning starts, on the argument that a poor opening constrains every operation that follows. A Critic renders the candidate, scores it and hands back a diagnosis tied to specific operations, parameters and timings. The motion score averages six criteria judged against the script, the opening-view score averages three, and a plan is accepted only above thresholds of 0.62 for motion and 0.70 for the opening.

The appendix case study shows why the opening gets its own agent. Given an instruction for a loose upper-body portrait that glides forward at a fixed focal length, the Director's own opening sat at 41.88 mm and cropped the top of the hair, with a fixed lens leaving no way to loosen the framing as the camera advanced. The Scout placed the camera higher and pitched it slightly down, at an elevation of -7.70 degrees, and chose a 26.49 mm lens, which showed the full hair silhouette and left room for the push-in.

The Director also draws on an example bank built only from CinemaBench's development scenes. Retrieval runs on up to three keywords that the Interpreter assigns from a shared vocabulary of visual goals, motion requirements and framing requirements, matched with Jaccard similarity, so an example transfers by intent rather than by subject. The paper's illustration retrieves a validated plan that dollies toward a monster while holding its framed size, then adapts that same three-operation combination, Dolly In with Frame Subject and Aim Follow, to a new shot pushing toward a bunny without letting it grow.

Cost per task averages 2.23 minutes and 69.5k tokens over 6.23 model calls, a figure that counts failed attempts and excludes the final video export. All four agents run on gpt-6-astra at medium reasoning effort, which makes FilmFlow an agent loop that calls a frontier model several times for one shot.

CinemaBench, the benchmark shipped with the system, is 150 Blender scenes, 73 dynamic and 77 static, built from self-created and licensed assets. Each scene carries five camera instructions, one easy, two medium and two hard, for 750 text-directed tasks, plus 60 reference clips averaging about 10 seconds, drawn from excerpts of publicly released films and from the authors' own footage. The split is by scene: 50 development scenes feed only the example bank, and the 100 test scenes carry every reported number, 500 text-directed and 40 reference-directed tasks. Of the 750 instructions, 50 were written by the authors and 700 were generated with gpt-6-astra, then filtered by programmatic checks and a model-based review. No ground-truth trajectory is published for the text-directed tasks, because several different operation sequences can satisfy one instruction. The benchmark judges rendered behavior instead of a match to a single answer key, which is the right call for camera direction and also means it measures preference more than correctness.

Scoring is handled by outside models: VideoReward for instruction alignment and motion quality, the DOVER composite for overall quality on a 0 to 100 scale, and a Probe that applies the CameraBench camera-motion model to report the probability that a clip matches its instruction. The paper states that none of those evaluators share weights, prompts or training data with FilmFlow, which stops the system from grading its own homework. On text-directed tasks FilmFlow scores 47.60 for quality against 44.65 for CinemaTraj, 41.54 for LAMP and 39.63 for GenDoP, and 0.555 for motion quality against a best baseline of 0.257. On reference-directed tasks it scores 41.77 for quality against 35.86 for a zero-shot VLM, 35.55 for LingBot-Map and 34.88 for VGGT-Omega, with a Probe score of 0.750 against 0.533, 0.521 and 0.502.

The ablation table is tighter than the headline. Removing the Scout, banning overlapping operation windows, disabling the Critic's revisions, or dropping the retrieved examples each costs the full system something, and full FilmFlow still wins every column except one: on the reference-directed Probe score the no-ICL variant reaches 0.766 against 0.750, which the paper attributes to Probe's auxiliary role. Read those Probe numbers as directional.

The human study is the stronger evidence. Forty-two participants each judged 10 text-directed and 10 reference-directed trials, with method identities hidden and presentation order randomized, producing 420 and 397 completed rankings. FilmFlow's pass rate was 84.3% on text-directed trials and 82.1% on reference-directed ones, with top-1 rates of 61.0% and 63.2%. For comparison, CinemaTraj passed 63.8% of text-directed judgments and took first place in 21.2% of them, while LingBot-Map passed 77.3% of reference-directed judgments and took first in 17.4%. On a five-point scale FilmFlow's mean ratings were 4.2, 4.4 and 4.3 for alignment, visual quality and motion smoothness on the text side.

Two things temper the comparison. FilmFlow's four agents and the zero-shot VLM baseline all run on gpt-6-astra at medium reasoning effort, so part of the contest is one orchestration recipe against another on a shared frontier model, and the baseline systems were adapted into the target scene's coordinate frame by the authors themselves. The paper also notes that GenDoP and LAMP do not track moving subjects, which it calls a limitation of those methods rather than of the adaptation. CinemaBench is Blender-only, and FilmFlow needs the scene to already exist: planning a camera inside a scene somebody built is the task, and the authors' listed next step is extending the system to automate scene construction as well.

Nothing is downloadable yet. The project page lists the code as a release link to be added and CinemaBench as coming soon, and the paper's reproducibility statement says the implementation and annotations will be released publicly along with instructions for acquiring the required scene assets. The ethics statement flags that some third-party assets may not be redistributable, so the release could arrive in parts.

What a creator can take from FilmFlow today is the camera-plan idea and one application. Describing a shot as named operations with explicit start and end times is a model any camera-control workflow can borrow, and the appendix shows why it matters downstream: gray-model renders of FilmFlow trajectories were used as structural conditions for Seedance video generation, with one camera plan reused across four prompts. Framing, subject placement and timing held while appearance followed the prompt. That decoupling, a camera plan that survives a change of art style, is the part of this paper that reaches a real pipeline soonest. It lands in a week of research-class releases that includes SLOP TV's own report on Peking University and Kling's artifact detector.

The paper is on arXiv as 2610.30753, the demo clips are on the project page, and the code and CinemaBench links still read coming soon, so you can watch FilmFlow's camera moves now but not run them.