{"slug": "filmworld-agentic-novel-to-film-generation", "title": "FilmWorld: Agentic Novel-to-Film Generation", "summary": "Researchers at Huazhong University of Science and Technology and Alibaba Group's Wan Team have developed FilmWorld, a system that generates films from novels by modeling a dynamic cinematic world as a five-tuple with persistent entity states and causal transitions. FilmWorld coordinates six agent groups across construction and evolution phases to achieve constant-latency rendering independent of shot count, outperforming five agentic baselines on 15 novels evaluated by the automated FilmEval protocol across nine metrics.", "body_md": "# Turning novels into films by\n\nmodeling a dynamic cinematic world\n\nFilmWorld 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.\n\n1\n\n**Huazhong University of Science and Technology**\n\n2\n\n**Wan Team, Alibaba Group**\n\n## See FilmWorld in action\n\nA 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.\n\n## A novel is not a longer clip — it is an evolving world\n\nVideo 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.\n\n#### Independent per-shot synthesis\n\nEach 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.\n\n#### Sequential extrapolation\n\nShots are rolled out one after another with limited context. Semantic drift compounds over long narratives, and latency grows linearly, O(N), with shot count.\n\n#### A persistent, evolving world\n\nWe 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.\n\n## Dynamic Cinematic World Modeling\n\nWe 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.\n\n**Entity Set** Characters, locations, and props inferred and instantiated from the prose.\n\n**State Space** Per-entity states: age, costume, emotion, weather, time of day, physical condition.\n\n**State Identifier** A deterministic hash addressing each state to a reusable visual asset.\n\n**Transition Function** Plot events drive causal state updates: W\n\nt+1= 𝒯(W\n\nt, p\n\nt).\n\n**Rendering Function** Synthesizes a shot from the active state: v\n\nt= ℛ(W\n\nt, d\n\nt).\n\nt→ W\n\nt+1\n\n**Character · Mathilde Loisel** state @ W₁\n\nA plot event 𝒯(·, pt) advances the state; Φ maps it to a stable visual identity before ℛ renders the shot.\n\n## Six specialized agent groups, two phases\n\nModeling a cinematic world spans heterogeneous skills. FilmWorld coordinates six agent groups across a Construction side and an Evolution side.\n\n#### Narrative Structured Translation\n\nSegments the novel, resolves aliases and pronouns into canonical entities, and recovers cinematic details prose leaves implicit.\n\n#### World Entity State Modeling with Visual Anchoring\n\nDiscretizes states, hashes them into identifiers Φ, and anchors each first appearance to a reusable reference asset.\n\n#### State-driven Shot Planning\n\nMaterializes the full state trajectory and shot directives, chaining each shot's opening to its predecessor's end-state.\n\n**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.\n\n## See it on all 15 novels\n\nPick any source novel, read its opening in the original language or a translation, then compare the films inferred by FilmWorld against five agentic baselines.\n\n### The Necklace\n\n## How one novel becomes a world\n\nReal assets from FilmWorld constructing *The Necklace*: entities are anchored to reference visuals, then states evolve into scene-level keyframes. Click any frame to enlarge.\n\n## FilmEval — nine metrics, three dimensions\n\n15 novels across easy, medium, and hard tiers, scored by an automated multimodal protocol that jointly reads the source novel and the generated film.\n\n#### Cinematic Presentation\n\nFilm-level production quality: is the output a coherent, watchable cinematic artifact?\n\n**Visual Presentation**— stability, plausibility, aesthetic appeal, style adherence.\n\n**Narrative Expression & Pacing**— clarity of story, transitions, and rhythm.\n\n**Audio-Visual Performance**— dialogue clarity, ambience, and synchronization.\n\n#### Film Consistency\n\nDo visual entities and environments stay coherent across shots and time?\n\n**Character Consistency**— identity stability and story-role alignment.\n\n**Scene Consistency**— stable, recognizable, separable locations.\n\n**Object Consistency**— persistence and plausibility of salient props.\n\n#### Novel Fidelity\n\nIs the film faithful to the source — no fabrication, sound logic, full coverage?\n\n**Hallucination Resistance**— avoids unsupported or conflicting content.\n\n**Logic Reliability**— preserves causal, temporal, and relational logic.\n\n**Story Restoration**— coverage of fine-grained source details.\n\n## State of the art, stable across difficulty\n\nUnder a unified foundation stack and a shared Ghibli-anime style, FilmWorld leads every evaluation cell — and its advantage widens most on the hardest novels.\n\n#### Overall FilmEval score\n\nSix annotators rated all 90 films. FilmWorld ranks first in every difficulty–dimension cell, and the human ranking matches FilmEval exactly (Spearman ρ = 1.0).\n\n| Method | Year | Easy | Medium | Hard | Overall | ||||||\n|---|---|---|---|---|---|---|---|---|---|---|---|\n| CP | FC | NF | CP | FC | NF | CP | FC | NF | |||\n| MM-StoryAgent | 2025 | 81.53 | 84.40 | 78.49 | 77.47 | 81.22 | 74.44 | 80.47 | 87.70 | 78.88 | 80.51 |\n| VGoT | 2025 | 83.27 | 86.93 | 76.82 | 80.33 | 87.72 | 82.37 | 80.00 | 87.99 | 72.74 | 82.02 |\n| MovieAgent | 2025 | 77.27 | 84.28 | 85.88 | 78.93 | 87.07 | 86.62 | 81.73 | 89.43 | 84.63 | 83.98 |\n| ViMax | 2026 | 83.13 | 87.66 | 78.48 | 76.27 | 85.94 | 78.61 | 78.73 | 87.60 | 73.55 | 81.11 |\n| VideoClaw | 2026 | 82.27 | 86.15 | 84.22 | 81.67 | 88.34 | 82.15 | 80.53 | 88.12 | 84.28 | 84.19 |\n| FilmWorld (Ours) | 2026 | 84.00 | 90.93 | 92.66 | 82.47 | 91.56 | 92.91 | 82.87 | 92.95 | 94.09 | 89.38 |\n\n## BibTeX\n\n```\n@article{zuo2026filmworld,\n  title={FilmWorld: Agentic Novel-to-Film Generation through Dynamic Cinematic World Modeling},\n  author={Zuo, Jialong and Zuo, Haotong and Zhang, Shiwei and Wang, Xiang and Li, Chen and Sang, Nong and Gao, Changxin and Bai, Xiang},\n  journal={arXiv preprint arXiv:2607.19038},\n  year={2026}\n}\n```\n\n", "url": "https://wpnews.pro/news/filmworld-agentic-novel-to-film-generation", "canonical_source": "https://filmworld-ai.github.io/", "published_at": "2026-07-22 03:46:51+00:00", "updated_at": "2026-07-22 04:22:43.305423+00:00", "lang": "en", "topics": ["artificial-intelligence", "generative-ai", "ai-agents", "computer-vision", "natural-language-processing"], "entities": ["Huazhong University of Science and Technology", "Alibaba Group", "Wan Team", "FilmWorld", "FilmEval"], "alternates": {"html": "https://wpnews.pro/news/filmworld-agentic-novel-to-film-generation", "markdown": "https://wpnews.pro/news/filmworld-agentic-novel-to-film-generation.md", "text": "https://wpnews.pro/news/filmworld-agentic-novel-to-film-generation.txt", "jsonld": "https://wpnews.pro/news/filmworld-agentic-novel-to-film-generation.jsonld"}}