Agentic Novel-to-Film Generation · 2026

Turning novels into films by
modeling a dynamic cinematic world

FilmWorld reframes narrative long-video generation as Dynamic Cinematic World Modeling — first constructing a stateful, persistent world from prose, then evolving it causally scene by scene.

One click, and a crew of AI agents turns your novel into a complete film.
Jialong Zuo1,2, Haotong Zuo1, Shiwei Zhang2, Xiang Wang2, Chen Li1, Nong Sang1, Changxin Gao1, Xiang Bai1
1 Huazhong University of Science and Technology 2 Wan Team, Alibaba Group
First Look

See FilmWorld in action

A quick look at films FilmWorld generates automatically from a short novel — here, 华强买瓜 · Huaqiang Buying Watermelons, rendered in a few different visual styles. Click any to watch.

Huaqiang buying watermelons — Ghibli anime render Ghibli · Anime
HAND-DRAWN · WARM
华强买瓜 Huaqiang Buying Watermelons Hand-drawn 2D animation · Studio-Ghibli warmth
Huaqiang buying watermelons — Disney 3D render Disney · 3D
VOLUMETRIC · BRIGHT
华强买瓜 Huaqiang Buying Watermelons Volumetric 3D · Pixar / Disney lighting
Huaqiang buying watermelons — live-action render Live-action · Cinema
PHOTOREAL · CINE
华强买瓜 Huaqiang Buying Watermelons Photorealistic · film-graded colour
0
Overall FilmEval score, best in every cell
+0
Points over the strongest baseline
0
Difficulty-graded source novels
0
Automatic metrics · 3 dimensions
O(1)
Evolution-phase latency in shot count
0
Human rating vs 3.33 for the best baseline
The Challenge

A novel is not a longer clip — it is an evolving world

Video models excel at short, single-scene clips. A film demands dozens of scenes where every character, location, and prop carries state forward. Two habits break down at this scale.

Paradigm 01

Independent per-shot synthesis

Each shot is generated alone with no cross-shot state. A character injured in scene 3 reappears unharmed in scene 9 — identity and continuity simply are not tracked.

Paradigm 02

Sequential extrapolation

Shots are rolled out one after another with limited context. Semantic drift compounds over long narratives, and latency grows linearly, O(N), with shot count.

FilmWorld

A persistent, evolving world

We build one cinematic world grounded in the novel and evolve it through plot-driven state transitions. Every shot is just a view rendered from that world at a moment in time.

Core Formulation

Dynamic Cinematic World Modeling

We formalize the cinematic world as a five-tuple. Every shot is a rendered view of this world; consistency becomes a structural guarantee rather than an emergent hope.

Entity SetCharacters, locations, and props inferred and instantiated from the prose.
𝒳
State SpacePer-entity states: age, costume, emotion, weather, time of day, physical condition.
Φ
State IdentifierA deterministic hash addressing each state to a reusable visual asset.
𝒯
Transition FunctionPlot events drive causal state updates: Wt+1 = 𝒯(Wt, pt).
Rendering FunctionSynthesizes a shot from the active state: vt = ℛ(Wt, dt).
World-state trajectory · Wt → Wt+1
W₁
W₂
W₃
W₄
W₈
W₇
W₆
W₅
Character · Mathilde Loiselstate @ W₁

A plot event 𝒯(·, pt) advances the state; Φ maps it to a stable visual identity before ℛ renders the shot.

The FilmWorld System

Six specialized agent groups, two phases

Modeling a cinematic world spans heterogeneous skills. FilmWorld coordinates six agent groups across a Construction side and an Evolution side.

GROUP 01

Narrative Structured Translation

Segments the novel, resolves aliases and pronouns into canonical entities, and recovers cinematic details prose leaves implicit.

GROUP 02

World Entity State Modeling with Visual Anchoring

Discretizes states, hashes them into identifiers Φ, and anchors each first appearance to a reusable reference asset.

GROUP 03

State-driven Shot Planning

Materializes the full state trajectory and shot directives, chaining each shot's opening to its predecessor's end-state.

O(N) → O(1)

Parallelism through explicit state externalization. Because the entire state trajectory is materialized symbolically before any pixel is drawn, rendering carries no sequential dependency — the runtime-dominant Evolution Phase becomes constant-latency in shot count, a Θ(N) theoretical speedup.

Inside a Cinematic World

How one novel becomes a world

Real assets from FilmWorld constructing The Necklace: entities are anchored to reference visuals, then states evolve into scene-level keyframes. Click any frame to enlarge.

The Benchmark

FilmEval — nine metrics, three dimensions

15 novels across easy, medium, and hard tiers, scored by an automated multimodal protocol that jointly reads the source novel and the generated film.

Cinematic Presentation

Film-level production quality: is the output a coherent, watchable cinematic artifact?

VPVisual Presentationstability, plausibility, aesthetic appeal, style adherence.
NEPNarrative Expression & Pacingclarity of story, transitions, and rhythm.
AVPAudio-Visual Performancedialogue clarity, ambience, and synchronization.

Film Consistency

Do visual entities and environments stay coherent across shots and time?

CCCharacter Consistencyidentity stability and story-role alignment.
SCScene Consistencystable, recognizable, separable locations.
OCObject Consistencypersistence and plausibility of salient props.

Novel Fidelity

Is the film faithful to the source — no fabrication, sound logic, full coverage?

NHRHallucination Resistanceavoids unsupported or conflicting content.
LRLogic Reliabilitypreserves causal, temporal, and relational logic.
SRStory Restorationcoverage of fine-grained source details.
Quantitative Results

State of the art, stable across difficulty

Under a unified foundation stack and a shared Ghibli-anime style, FilmWorld leads every evaluation cell — and its advantage widens most on the hardest novels.

Overall FilmEval score

Human evaluation · 540 judgments
4.25/5vs3.33

Six annotators rated all 90 films. FilmWorld ranks first in every difficulty–dimension cell, and the human ranking matches FilmEval exactly (Spearman ρ = 1.0).

MethodYearEasyMediumHardOverall
CPFCNFCPFCNFCPFCNF
MM-StoryAgent202581.5384.4078.4977.4781.2274.4480.4787.7078.8880.51
VGoT202583.2786.9376.8280.3387.7282.3780.0087.9972.7482.02
MovieAgent202577.2784.2885.8878.9387.0786.6281.7389.4384.6383.98
ViMax202683.1387.6678.4876.2785.9478.6178.7387.6073.5581.11
VideoClaw202682.2786.1584.2281.6788.3482.1580.5388.1284.2884.19
FilmWorld (Ours)202684.0090.9392.6682.4791.5692.9182.8792.9594.0989.38

CP · Cinematic Presentation · FC · Film Consistency · NF · Novel Fidelity. FilmWorld stays within a tight ~1-point band across tiers (89.2 → 89.0 → 90.0), while baselines drop by up to 3.2 points as difficulty rises.

Citation

BibTeX

@article{zuo2026filmworld,
  title={FilmWorld: Agentic Novel-to-Film Generation through Dynamic Cinematic World Modeling},
  author={Zuo, Jialong and Zuo, Haotong and Zhang, Shiwei and Wang, Xiang and Li, Chen and Sang, Nong and Gao, Changxin and Bai, Xiang},
  journal={arXiv preprint arXiv:2607.19038},
  year={2026}
}
Global Reach
Locating you on the map…
Total page views