DyCraft breaks long-horizon 3D animation into storyboard steps for coding agents
Original title: DyCraft: An Agentic System for Long-Horizon 4D Scene Generation
On 24 multi-shot tasks, human raters preferred DyCraft over a native GPT-6 agent in Codex and prior baselines.
The paper introduces DyCraft, an agentic system for long-horizon animated 3D scene generation. It expands a prompt into a plan, builds a static scene, then a 3D storyboard of key poses so agents can debug frames and author local motion with motion tools. Bare coding agents often produce interpenetrating, unsupported, or unnatural motion by scripting the full animation in one pass. The authors report fewer physical violations and higher-quality scenes than a native GPT-6 agent in Codex and prior agentic baselines.
Unconfirmed: Quality and preference gains are reported by the paper's own evaluation, not an independent study.
- Evaluation covers 24 long-horizon, multi-shot animation tasks.
- Tasks include humanoid robots, human and animal characters, and physical mechanisms.
- Storyboards turn animation into pairs of verified start and end states.
- Authors say DyCraft cuts physical violations versus a GPT-6 agent in Codex.
- Products
- Codex
From the original article
1 min read at huggingface.co
While frontier coding agents can generate high-quality 3D scenes and articulated assets, they struggle to create coherent long-horizon animations. When equipped with general-purpose harnesses, they frequently produce physically implausible outputs, generating objects that interpenetrate, lack physical support, or move unnaturally.
By Jeff Tan, Hang Dai, Zeqian Long · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.