CausalCine: Causality in Film & Video

• CausalCine is a dual-use framework that extracts causal event pairs from film scene descriptions and enables real-time multi-shot video narrative generation.
• The film-event approach leverages screenplay text with techniques like Combined CP and human evaluation to quantify narrative causality across physical, motivational, psychological, and enabling types.
• The video-generation method uses an interactive autoregressive framework with content-aware memory routing to maintain coherence, prompt updates, and shot-boundary integrity in long video sequences.

CausalCine denotes two distinct research usages centered on filmic sequence structure. In one usage, it is a system for extracting and ranking causal relations between events in film scene descriptions, drawing primarily on Hu and Walker’s work on narrative causality in screenplay corpora (Hu et al., 2017). In the other, it is an interactive autoregressive framework for real-time multi-shot video narrative generation that generates causally across shot changes, accepts dynamic prompts on the fly, and reuses context without regenerating previous shots (Meng et al., 12 May 2026). Across both usages, the organizing concern is temporally ordered cinematic material: scene descriptions whose temporal order approximates narrative order, and multi-shot sequences whose coherence depends on event progression, viewpoint shifts, and shot boundaries.

1. Dual usage of the term

The supplied literature uses the name for two different technical programs.

Usage	Primary task	Core substrate
Film-event CausalCine	Extract and rank causal event pairs	Film scene descriptions
Video-generation CausalCine	Real-time autoregressive multi-shot generation	Multi-shot video sequences

The earlier usage concerns narrative understanding. It learns pairs of causal events from a corpus of film scene descriptions that are action rich and tend to be told in chronological order. The later usage concerns generation. It transforms multi-shot video generation into an online directing process and targets streaming interactivity under causal autoregression (Hu et al., 2017); (Meng et al., 12 May 2026).

A common source of confusion is that the shared name does not identify a single continuous method. One line of work infers causal relations between symbolic events; the other learns to generate video causally across shot changes. This suggests that the shared label is best understood as a recurring focus on cinematic causality and coherence rather than as a single architecture or dataset lineage.

2. Narrative causality as defined over film events

In the film-event setting, narrative causality is defined using narrative theories associated with Trabasso and van den Broek and with van den Broek. Four distinct causality types are distinguished: physical causality, motivational causality, psychological causality, and enabling causality (Hu et al., 2017).

Physical causality is the case in which event $A$ physically brings about event $B$ by direct force or contact. The examples given are grab(person, sleeve) → spill(person, beer) and push(person, foe) → stumble(person).

Motivational causality is the case in which $A$ occurs because the actor intends or is motivated to bring about $B$. The examples given are stoop(person) → avoid(person, hit_head) and bend(person) → pick_up(person, something).

Psychological causality is the case in which $A$ evokes or changes an emotional state expressed in $B$. The examples given are look(person, around) → enjoy(person, familiarity) and see(person, ghost) → fear(person).

Enabling causality is the case in which $A$ creates conditions or a pre-requisite state that makes $B$ possible. The examples given are pull(person, ring) → slide_off(person, ring) and clear_path(person) → proceed(person).

This typology matters because previous work on learning narrative relations from text had either focused on “strict” physical causality or had been vague about what relation was being learned. The film-event formulation therefore broadens causal inference beyond direct physical effects to include intention, affect, and precondition structure.

3. Corpus construction, event representation, and ranking functions

The film-event system draws on 955 films scraped from IMSDb, covering 11 genres, with total word count per genre ranging from approximately $1.1\text{M}$ in Fantasy up to approximately $6.7\text{M}$ in Drama. Scenes are filtered to remove dialogue so that only descriptive text remains. The motivation for using screen-play scene descriptions is twofold: they are action rich, and they recount events in chronological order, so temporal order provides a strong cue for causality (Hu et al., 2017).

Preprocessing uses Stanford CoreNLP for tokenization, lemmatization, POS tagging, parsing, and NER. Events are identified as all verb lemmas with POS VB*, excluding light verbs such as be, have, do, let, start, and try. For each event, the representation records its subject (nsubj, agent), object (dobj, nsubjpass), indirect object (iobj), and any verb particle (compound:prt).

Argument generalization maps entity or noun arguments to either person or something. An argument is mapped to person if it is a person named entity, a non-it pronoun, or a WordNet noun.person; otherwise it is mapped to something. This generalization is central to the induced event-pair representation because judgments are collected over generalized arguments rather than over film-specific entities.

Causality is quantified with Causal Potential (CP). For a window size $B$0,

$B$1

with

$B$2

To combine evidence across window sizes $B$3, the system defines Combined CP (CPC):

$B$4

This weighting rewards closely adjacent pairs more heavily than distant ones, and all event-pair frequencies are smoothed with add-one for unseen pairs. High-CPC pairs have both high PMI and strong directionality, whereas low-CPC pairs tend to be weakly associated or symmetric or sequential without causal asymmetry.

4. Human evaluation, causal typing, and stated limitations of event-pair induction

The film-event system is evaluated against low-scoring alternatives and against an external baseline. In the first Mechanical Turk experiment, 960 top-ranked high-CPC pairs are compared with 960 low-CPC pairs. Annotators are asked which pair is more likely to manifest true narrative causality. High-CPC pairs are chosen in 82.8% of cases by majority vote; 51% are unanimous among 5 annotators; and Krippendorff’s $B$5 is approximately $B$6 (Hu et al., 2017).

In the second experiment, CPC pairs are compared with Rel-grams pairs. Rel-grams use symmetric conditional probability,

$B$7

over newswire. On 100 items, CPC wins 81% versus Rel-grams, with $B$8 approximately $B$9.

The third experiment asks annotators to assign one of the four narrative-causality types to 100 high-CPC pairs. A majority agreement is reached for 79% of items. The reported distribution is Physical 13%, Motivational 29%, Psychological 9%, Enabling 28%, with others described as noisy. These figures indicate that the induced pairs are not confined to physical causality.

The fourth experiment compares pairs learned on genre subcorpora with pairs learned on the full 955-film corpus. There is over 70% overlap among the top 960 pairs, and there is no significant difference in judged quality, with an approximately 50–50 split on unique pairs. Genre-specific lists nevertheless yield distinctive pairs, including listen→muffle for horror and clear→reveal for science fiction.

The limitations stated for this system are specific. The current model only scores dyadic pairs rather than longer event chains. Proposed extensions include learning full causal chains or “scripts” by chaining high-CPC pairs and enforcing transitive constraints; cross-film generalization through entity linking or normalization of proper names; multimodal integration with video frames or audio cues; automatic classification of causality type with a supervised classifier using lexical patterns or semantic roles; incorporation of nominal events and adjectives; and the use of discourse structure such as RST or coreference to filter out spurious sequential pairs.

5. Causal autoregression for multi-shot video narratives

In the 2026 usage, CausalCine is an interactive autoregressive framework for multi-shot video generation. The motivating claim is that cinematic storytelling is not merely the endless extension of a single scene; it requires progressing through evolving events, viewpoint shifts, and discrete shot boundaries. Standard autoregressive video models are reported to struggle with two failure modes during long rollouts: motion stagnation and looping, and semantic drift (Meng et al., 12 May 2026).

Multi-shot narratives add further constraints. At a shot boundary, the prompt changes, so the model must abandon previous-shot dynamics and faithfully depict the new shot. Cross-shot identity preservation is also required, because a character may vanish for many seconds and then re-appear. Existing autoregressive video diffusion models are described as treating the entire sequence as one long shot, with no mechanism to detect or respect shot boundaries, inject new prompts at runtime, or content-addressably recall frames across many seconds.

The causal base model starts from a pretrained bidirectional video diffusion transformer in latent space. Given a clean video latent $A$0 and noise $A$1, the noised latent is formed as

$A$2

and the velocity field $A$3 is trained with the rectified flow-matching loss

$A$4

For autoregression, a long multi-shot video is split into $A$5 contiguous chunks $A$6, each containing $A$7 latent frames, with $A$8, corresponding to approximately 12 video frames. Shot prompt changes are observed at chunk boundaries, and the factorization is

$A$9

where $B$0 is the text embedding of the shot prompt for chunk $B$1.

Training uses a “2N-segment packing” teacher-forcing scheme rather than naive step-by-step rollout. Each noisy query chunk attends causally to preceding clean context together with local window and long memory, and the resulting single-pass layout is stated to expose the same causal visibility patterns used at inference with a KV cache.

6. Content-aware memory routing, distillation, and interactive inference

The principal memory mechanism is Content-Aware Memory Routing (CAMR). At inference time, the model maintains a growing set of key/value pairs for all past latent frames. To keep memory bounded, it combines a small sliding window $B$2 of the $B$3 most recent chunks with a fixed-size content-addressable buffer of $B$4 frames drawn from the distant past (Meng et al., 12 May 2026).

Retrieval is descriptor-based. For cached keys $B$5, the method computes a frame descriptor

$B$6

and a mean-pooled query descriptor $B$7 for the current chunk. Semantic relevance is then scored by

$B$8

The top-$B$9 frames are retrieved from the out-of-window history and combined with the local window and the current chunk. Because retrieved frames may lie far beyond the absolute positions seen in training, the method applies block-relative RoPE by re-anchoring all positions within a compact block range before applying 3D RoPE.

The full causal base model is then distilled into a four-step generator. The first stage is causal ODE initialization, which aligns a four-step student $A$0 with teacher PF-ODE trajectories under teacher forcing. The second stage switches to self-forced DMD, rolls out the student autoregressively, perturbs to $A$1, and applies the implicit DMD gradient. To stabilize long-sequence framing and camera motion, the student is augmented with a lightweight GAN head $A$2, and the final generator loss is

$A$3

Interactive generation is organized chunk by chunk using the KV cache and CAMR. New shot prompts may be appended on the fly, and chunks beyond a shot boundary automatically switch their cross-attention keys to the new prompt. Previously generated chunks are never re-computed; their KV pairs are cached and autoregression continues from the last chunk. Because CAMR retrieves historical frames by semantic relevance rather than temporal proximity, a character disappearing for tens of seconds can still be recalled when relevant.

7. Experimental profile, comparative results, and conceptual relation between the two usages

The multi-shot video system is benchmarked on a 100-prompt multi-shot suite with 5 shots each and a total duration of approximately 15 seconds. The reported metrics are Aesthetic score (LAION), Text–video alignment (ViCLIP similarity), Subject and background consistency (DINO, CLIP), and Shot-cut accuracy (SCA) (Meng et al., 12 May 2026).

Against autoregressive baselines—Self-Forcing, Infinity-RoPE, LongLive, MemFlow, and ShotStream—CausalCine achieves the highest aesthetic quality, better prompt alignment, and the best subject/background consistency and SCA. The specific examples reported are ViCLIP $A$4 versus best autoregressive baseline $A$5, and SCA $A$6 versus second best $A$7.

Compared with bidirectional multi-shot methods—HoloCine and MultiShotMaster—the four-step causal model matches or exceeds their text alignment and intra-shot consistency, while running more than $A$8 faster in streaming mode and supporting interactive prompt updates. Ablation results further report that training a native causal multi-shot base model before distillation is critical, because directly compressing a short-clip bidirectional teacher yields severe cross-shot drift. They also report that content-aware memory routing outperforms no-memory or first-frame sink strategies in preserving distant-shot identities, with inter-shot consistency increasing to $A$9 from $B$0 and $B$1.

Placed together, the two usages of CausalCine mark different computational interpretations of causality in cinema. One induces narrative causal relations between event pairs from action-rich, chronologically ordered scene descriptions; the other performs causal generation over chunked multi-shot video with bounded memory and prompt updates. This suggests a broad thematic continuity—causal organization in film narrative—even though the concrete objects of inference, the learning objectives, and the evaluation protocols are entirely different (Hu et al., 2017); (Meng et al., 12 May 2026).