Remediating Alfred Hitchcock: From Parametric to Recursive Montage

TL;DR

A film-theory / media-archaeology essay (Bibliotheca Hertziana – Max Planck Institute for Art History) that situates contemporary text-to-video and “world-simulating” generators (Sora, Runway Gen-1, FLUX.1 Kontext, Wan 2.6) inside a 70-year intermedial history of montage. It argues the cinematic operation is moving from the splice (Eisenstein) through parametric (rule-governed, computer-assisted: John Whitney’s analog-computer animation for Saul Bass’s Vertigo opening titles) and exhibited-cinema (Douglas Gordon, Godard, Binotto) modalities into a recursive configuration where multi-model chains (ComfyUI orchestrating LLaVA-Phi + FLUX.1 Kontext + Wan 2.6) generate self-referential, in-principle infinite films from latent-space rollouts. The conclusion is the part that matters for the Luma audience: a cautionary read of “world models” as fundamentally cinematic artifacts, and an explicit warning that the sim-to-real gap is structural rather than a quality-of-fidelity problem.

Key claims

Generative video models do not assemble recorded footage but synthesize temporal relations from within through rule-governed and feedback-driven processes — a categorical shift from the classical cut to “parametric” and ultimately “recursive” montage, anticipated by Gene Youngblood in 1989 [§Introduction].
The author maps three historical phases: (i) parametric montage in John Whitney’s 1957 Lissajous-curve animations for Vertigo’s opening titles, made by repurposing an M5 anti-aircraft gun director (a WWII servo-mechanism with continuous feedback correction) as a drawing stand [§1]; (ii) exhibited cinema (Douglas Gordon’s 24 Hour Psycho, Godard’s Histoire(s) du cinéma, Binotto’s Metaleptic Attack) that treats cinema as rewritable material via VHS/DV iterative manipulation [§2]; (iii) recursive montage in contemporary generative-AI pipelines [§3].
Recursive montage is operationalized concretely as Grégory Chatonsky’s Readonlymemories VI (2026): a ComfyUI-orchestrated chain — LLaVA-Phi for image-to-text captioning → FLUX.1 Kontext for single-image generation → Wan 2.6 UHD for high-resolution video synthesis — where each new 5-second segment is predicted from the textual description of the preceding one, producing a self-referential feedback loop that extends Vertigo into “a computationally mediated, in principle infinite film” [§3].
The conclusion frames world-model deployments for robotics and embodied agents as a category error: gaming/entertainment applications are legitimate (films/games are themselves “limited representational systems”), but extending models trained on mediated representations to encounter the physical world is fundamentally unjustified — “driving a car is not like racing in Forza Horizon, cleaning a house is not like playing The Sims, and fighting in war is not like shooting in Call of Duty: Zombies” [§Conclusion].
The argument extends Hollis Frampton’s 1971 notion of the “infinite film” — a database whose vault encompasses all possible cinematic material — to read latent-space rollouts as a contemporary instantiation of that idea, while distinguishing it from physical reality [§Conclusion].

Method

A historical/critical reading rather than an empirical study. The author traces a media-archaeological lineage through three case studies — Whitney’s analog-computer collaboration with Saul Bass on Vertigo (1958); Gordon, Godard, and Binotto’s VHS/DV-era essayistic remediation of Hitchcock; and Chatonsky’s text-to-image / video-generator series After the Cinema (1997–2026) — using David Bordwell’s 1986 concept of parametric narration (“style supersedes representational necessity”) and Hollis Frampton’s 1971 “infinite film” as the load-bearing theoretical framework. The closing section reads contemporary multi-model pipelines (ComfyUI + LLaVA-Phi + FLUX.1 Kontext + Wan 2.6) as instances of recursive montage — feedback-driven generation operating within learned latent spaces — and uses this framing to deliver a critique of the deployment of such systems as world models for physical-AI applications.

Results

No quantitative results — this is a theory paper. The substantive deliverables are (a) a vocabulary (classical cut → parametric → recursive montage) for reading generative video models as media artifacts with a 70-year prehistory; (b) a concrete pipeline diagram of Readonlymemories VI as the first artifact instance of “recursive cinematic montage”; (c) a normative position that world models occupy a long lineage of cinematic operations rather than world-simulating ones, with strong implications for how the embodied-AI community should interpret their outputs.

Why it’s interesting

This is a rare wiki entry sourced from outside ML/CS — a media-studies framing of the same world-model artifacts the rest of the cluster treats as engineering objects, and a sharply different critical register on the sim-to-real question. It sits in productive tension with The flavor of the bitter lesson for computer vision (which argues video-generative pre-training should replace explicit 3D as the substrate for embodied AI) by inverting the framing: where Sitzmann reads video models as the right pre-training objective, Boutet de Monvel reads them as fundamentally cinematic and therefore structurally insufficient for physical-world deployment. The argument also complements A Functional Taxonomy of World Models‘s renderer/simulator/planner split — Boutet de Monvel’s “recursive montage” is roughly Fei-Fei Li’s “renderer” category, and the essay supplies a media-archaeological reason why the renderer/simulator boundary matters. Empirical follow-ups in the wiki bear on the same question quantitatively: VideoPhy-2: A Challenging Action-Centric Physical Commonsense Evaluation in Video Generation (no model beats 33% on physical commonsense), RISE-Video: Can Video Generators Decode Implicit World Rules? (22.5% strict-accuracy ceiling on world-rule reasoning), and Dream.exe: Can Video Generation Models Dream Executable Robot Manipulation? (visual quality fails to predict executability ranking) all produce numerical versions of the gap this essay names from theory.