Skip to content
Bud's Wiki

Moonvalley Marey — controllable AI video model with 360° camera, pose, and motion transfer

TL;DR

Section titled “TL;DR”

Moonvalley launches Marey, a closed-source generative video model pitched as “the most controllable AI video model ever built” and aimed at film production. Marketed control surface: 360° camera control, high-definition pose mapping (skeleton transfer from human video), motion transfer from a driving video, and explicit per-object/character motion-and-trajectory control. The model is positioned as the first frontier video model “trained exclusively on fully-licensed content” — i.e. commercially safe for studios. Public-access launch announced July 8, 2025; no paper, weights, architecture, or benchmarks disclosed.

Key claims

Section titled “Key claims”
  • Marey advertises four control axes — 360° camera control, pose mapping, motion transfer from a driving video, and per-object/character trajectory control — in a single model [LinkedIn post].
  • The model is the “first ever trained exclusively on fully-licensed content,” positioning it as commercially safe for professional production [LinkedIn post].
  • Built by Moonvalley with Oscar-nominated filmmaker input; framed as “hybrid film-making” rather than prompt-only generation [LinkedIn post].
  • Available for public use via a Moonvalley link at launch [LinkedIn post].

Method

Section titled “Method”

Not disclosed. The announcement is a marketing post and links to a redirect (lnkd.in/eEh-VyzY) rather than a technical report. No architecture, parameter count, training-corpus size, video length / resolution targets, or evaluation numbers are given. The “360° camera control” claim is the most specific technical statement and implies some form of camera-trajectory conditioning (either explicit pose-sequence input à la ReCamMaster / VerseCrafter, or reference-video transfer à la OmniTransfer — the post doesn’t specify which).

Results

Section titled “Results”

No quantitative results. No comparisons with frontier closed models (Veo, Sora, Gen-4, Kling) or with open camera-controlled stacks. The only signal is the marketing copy and the launch availability.

Why it’s interesting

Section titled “Why it’s interesting”

Marey extends the small set of closed-flagship, productized video-generation efforts the wiki tracks under World Foundation Models — alongside Genie 3 / Project Genie (DeepMind), the Waymo World Model post-trained from Genie 3, and Runway Labs’s portfolio-incubator framing. Of these, Marey is the only one foregrounding explicit multi-axis film-production controls (camera + pose + motion + object) as the differentiator, rather than open-ended interactive rollout (Genie) or vertical specialization (Waymo). It is also a fresh datapoint for Camera-Controlled Video Diffusion: every filed paper in that cluster is open and built on a frozen Wan2.x backbone with adapters, while Marey is closed and trained from scratch on licensed data — so it cannot be compared head-to-head with Kling-MotionControl, OmniTransfer, or Wan-Animate, but it is the clearest non-open instance of the same “stack camera + pose + character + motion control on one video model” thesis. The licensed-data claim, if real, is the more unusual axis and worth tracking against later legal/IP disclosures.

See also

Section titled “See also”