Moonvalley gives filmmakers reference-driven control over AI video

Moonvalley develops Marey, a video-generation model aimed at filmmakers who need to direct a shot rather than accept whatever a short prompt produces. Its offer combines text and image generation with controls informed by reference material. The company also emphasizes licensed training data. Those are meaningful evaluation topics, but neither a control feature nor a licensing claim establishes that a particular output is ready for a client’s production.

01 / ProductMarey connects generative video to directorial inputs

The Marey overview presents camera control, motion transfer, trajectory control, keyframing and reference inputs as ways to shape a generated shot. These controls address different questions. A reference image describes appearance, a motion reference describes movement, and a camera instruction describes how the scene is viewed. A buyer should test the control that solves the actual production problem.

Moonvalley says Marey is trained on licensed footage. That is a vendor statement about the training collection, not our independent audit of every training record or a guarantee about all possible outputs. It may be relevant to a studio’s selection criteria, but input permissions and the proposed use still matter.

The company’s fal launch announcement identifies Marey Realism v1.5 and links text-to-video, image-to-video, motion-transfer and pose-transfer endpoints. This gives developers a documented partner route. The current website also links a web application and labels its direct API route as a waitlist; those access paths should not be silently treated as the same service.

02 / AudienceFor shots that benefit from explicit reference material

A filmmaker who can supply a deliberate motion reference has a clearer starting point than someone asking for an entire finished film. A production designer can explore a setting, while an editor can assess whether a generated insert fits between two existing shots. Marey is most usefully evaluated against such bounded tasks.

An exact documentary record is a different requirement. If an image must prove what happened at a real event, generative footage cannot serve that purpose. For a fictional or clearly illustrative sequence, visual invention may be welcome. Decide which kind of material the audience will believe it is seeing and keep that distinction visible through production.

Teams also need someone who can judge temporal quality. A good thumbnail says little about motion through a whole clip. Review object continuity, contact with the ground, camera behavior and the transition into neighboring footage. When those decisions are central to the job, a director or experienced editor should help define the trial.

03 / WorkflowA proposed motion-transfer test for a short insert

For a proposed pilot, use a simple studio shot of an unbranded object moving across a table. Record or license a reference with the intended timing and camera movement. Choose a new fictional subject and prepare a first-frame image if visual continuity requires one. The aim is to test whether movement and appearance can be controlled separately.

The motion-transfer endpoint accepts a reference video and prompt, with an optional first-frame image, negative prompt and seed. Start with a reference whose movement is easy to describe. Avoid a busy montage: several simultaneous actions make it difficult to identify which part of the control failed.

Write a short shot specification covering framing, direction of travel, desired pace and the detail that must remain stable. Generate a candidate and compare it with the reference at normal playback speed. Then inspect selected frames around contact, occlusion and changes in direction. These are proposed checks, not claims that the tool will preserve every detail.

Run a second candidate with one changed requirement, such as a different surface material. Keep the other inputs stable. A controlled revision reveals whether the workflow can preserve a composition after approval. If the motion or camera changes unexpectedly, include the extra attempts and finishing effort in the evaluation.

The text-to-video endpoint offers a separate starting point with duration, dimensions, negative prompt and seed controls. Use it for a shot where the production is comfortable letting the model invent more of the scene. Do not assume a seed makes results identical across model revisions or access providers.

Export the chosen result into the actual editing timeline. Check cadence, crop and color beside the surrounding footage, then ask for a client-style revision. Keep the input reference, prompt, model identifier and accepted file together. The final evidence should explain how the shot was obtained and what still required manual work.

04 / PricingPartner API prices are not a web-app subscription

On 28 September 2026, the fal text-to-video page displayed US$1.50 for a five-second generation and US$3.00 for ten seconds. The motion-transfer page displayed US$2.00 per video. These are the partner’s stated endpoint charges, not a quotation for Moonvalley’s own application or an enterprise production agreement.

The Moonvalley website’s Pricing link returned its homepage during research, so a current direct subscription tariff was not established. Use the application or sales route to confirm its allowance and commitments. The absence of a readable tariff is not evidence that the web application is free.

As illustrative arithmetic, ten five-second text-to-video attempts at the displayed fal rate cost US$15 in endpoint charges. That is not US$15 for an approved shot: it excludes any additional processing, editing and review, and it assumes those attempts use the stated endpoint configuration. Track accepted footage separately from generated footage.

05 / DistinctionsReference control changes what a comparison should measure

Marey’s reference-led approach encourages a production team to communicate through visual and motion inputs. That can be more concrete than repeatedly rewriting adjectives in a prompt. The useful distinction is how much of an approved creative decision survives the next operation, not how many control names appear on a product page.

Runway is a relevant comparison for teams exploring both new footage and transformations of existing shots. Put the same narrow brief through each relevant workflow and compare the accepted result plus revision effort. A sample selected for a homepage is not a controlled comparison.

Luma AI provides another approach to generative video and visual control. The choice should reflect the reference material, output requirements and production environment available to your team. If the job requires a particular camera move, test that move instead of substituting an easier showcase scene.

06 / QuestionsWhere licensed-data claims meet a real production

The terms of service identify Moonvalley AI Inc. and address commercial outputs, authorized users and distribution partners. Partner access can involve additional terms. A studio should identify which agreement governs its route and whether any negotiated protection applies to the actual use, rather than reading “commercially safe” as an unrestricted guarantee.

Reference material can introduce its own obligations. A licensed model does not supply permission to upload somebody else’s performance, confidential footage or recognizable likeness. For the proposed pilot, use a reference the team controls and an unbranded subject. This makes the creative test easier to interpret as well as reducing uncertainty about inputs.

Check the endpoint and account before promising access to a control shown elsewhere on the website. A feature marketed for the application may not have the same parameters, availability or limits through a partner. The present research establishes public product descriptions and listed access routes; it does not establish production latency or contractual support response times.

Finally, retain a conventional route for a shot that repeatedly fails a fixed requirement. The right conclusion can be that generation is useful for previsualization while the final insert needs filming or compositing. That is a valid production decision, not a failed experiment.

07 / DecisionMake the next step a directed shot

Moonvalley is worth evaluating when a team has a specific shot, useful reference material and a reason to care about the model’s training-data position. Start with a controlled sequence and test one revision. Commit to a broader workflow only after the selected access route, costs and actual exported result meet the production’s needs.