Camera Artist: A Multi-Agent Framework for Cinematic Language Storytelling Video Generation
Haobo Hu, Qi Mao, Yuanhang Li, Libiao Jin
cs.AI
2026-04-10
Camera Artist adds recursive storyboards and a 580-pair cinematography LoRA on MAGREF, scoring 4.28 script consistency and 80 dynamic degree over MovieAgent and VGoT.
Clip-level text-to-video and image-to-video models already look good in isolation. Multi-agent pipelines such as MovieAgent, Anim-Director, and VideoGen-of-Thought can walk a script all the way to a stitched film. Adjacent shots still drift. Shot descriptions are usually written from the scene, with little conditioning on the previous shot, so characters and locations jump. Generic LLM prompts also skip professional lens language: shot size, camera move, framing, lighting. The result is a slideshow of plausible frames, not cinema.
Camera Artist, from Communication University of China and Peking University, treats those two gaps as the actual problem: how shots connect, and how each shot is photographed.
Three agents, two stages. Footage construction first, shot generation second.
The Director Agent (Qwen3-30B-A3B) expands a user outline into script, scenes, and character sheets via chain-of-thought, then calls Flux for character and scene reference images.
The Cinematography Shot Agent is the new piece. Recursive Shot Generation writes storyboards scene by scene. The opening shot of a scene sees only that scene and the global script. Every later shot is conditioned on the previous shot. The model itself decides when the scene has reached an endpoint. Cinematic Language Injection then rewrites each ordinary shot description into a film-style prompt. That rewriter is Qwen3-4B with LoRA (rank 8, alpha 32, 20 epochs) on 580 pairs from ShotBench. A VLM captions the raw clip without cinematic intent; GPT-4o folds ShotBench annotations (size, angle, motion, lighting) into the target sentence.
The Video Generation Agent runs MAGREF, a multi-reference image-to-video model, at 832×480 and 15 fps, then concatenates clips. The same pipeline can start from text only, with no character references.
Evaluation uses MoviePrompts (plots and character sheets from ten professional films) plus eight extra stories in the same format.
| Method | CLIP-T | Subject | Background | Motion | Dynamic | Aesthetic |
| VGoT | 28.15 | 78.58 | 97.93 | 99.27 | 16.67 | 68.73 |
| Anim-Director | 23.86 | 67.79 | 94.15 | 96.54 | 39.78 | 67.24 |
| MovieAgent | 22.25 | 71.01 | 94.52 | 98.00 | 76.27 | 65.63 |
| Camera Artist | 29.61 | 79.54 | 96.26 | 99.32 | 80.00 | 69.51 |
VGoT wins background consistency and almost freezes the camera (dynamic 16.67). Camera Artist posts the highest CLIP-T, dynamic degree (80.00), and aesthetic score.
Three VLMs (GPT-4o, Qwen3, Gemini-3) score 1-5. Averages: script consistency 3.90 vs best baseline 2.98; camera-movement consistency 3.55 vs 2.30; video quality 4.71; real-movie similarity 4.02. A five-point user study reports 4.28 on script consistency and 4.12 on real-movie similarity. Headcount is not given.
Ablations line up with the qualitative claims. Drop RSG and script consistency falls from 3.90 to 3.55, with sudden protagonist swaps. Drop CLI and camera-movement consistency falls from 3.55 to 2.83; shots go static.
For anyone building an AI short-film or storyboard stack, the paper splits "feels like a movie" into two plug-in modules: cross-shot conditioning, and a cinematography rewriter. The pixels still come from MAGREF and Flux. A 580-pair LoRA is enough to move the camera-movement score, which is a hint that lens language is still mostly prompt craft plus a small adapter, not a new video backbone.
This is an orchestration improvement. The quality ceiling is the underlying image-to-video model.
There is no dedicated limitations section. The test set is ten plus eight stories, and qualitative figures lean on Frozen and Zootopia characters. The user study does not report how many people scored. One of the three VLM judges is Qwen3, and so is the agent backbone, so family bias is untested. Recursion stays inside a scene; scene-to-scene joins are unmodeled. The "film" is concatenated clips, not a single long generation. CLI sits on 580 pairs, a 4B model, and LoRA rank 8; transfer beyond Hollywood shot grammar is not measured. Code and weights are not promised in the main text.