Video-Coder

Introduction

Video-Coder is a coding agent that writes and runs Blender code to build a draft video. A video generator then turns the draft into the final video.

+24VBench-2.0 over MiniMax-H3

+6PhyGenBench over MiniMax-H3

6k + 2kopen SFT / RL samples

Direct video generation vs. Video-Coder — **Figure 1:** Overview of our proposed Video-Coder. Unlike direct video generation, Video-Coder writes and executes code to build a 3D scene and renders a draft video, which guides the downstream video generator with explicit structure and motion signals.

Method

The agent works in the Video-Coder Harness, an environment where it writes, runs, inspects and revises its program. We train Qwen3.5-9B with SFT, then agentic RL with an execution reward that anneals into a video reward.

Figure 2: Overview of our data construction pipeline. It consists of four stages: text prompt construction, agentic trajectory generation, ground-truth video synthesis, and data cleaning and filtering, yielding two datasets: Video-Coder-SFT-6k and Video-Coder-RL-2k.

Results

Experiments show that Video-Coder consistently improves video generation performance across different backbones, bringing around 7 points gains for MiniMax-H3 on PhyGenBench, demonstrating that explicitly constructing 3D scenes with executable code effectively enhances the physical plausibility and controllability of generated videos.

Table 1: Performance of different models on the PhyGenBench benchmark. Best performance is in **bold**.
Method	Mechanics (↑)	Optics (↑)	Thermal (↑)	Material (↑)	Average (↑)
Pika	0.35	0.56	0.43	0.39	0.44
Gen-3	0.45	0.57	0.49	0.51	0.51
Kling	0.45	0.58	0.50	0.40	0.49
CogVideoX	0.39	0.55	0.40	0.42	0.45
Open-Sora V1.2	0.43	0.50	0.44	0.37	0.44
LaVie	0.30	0.44	0.38	0.32	0.36
Vchitect-2.0	0.41	0.56	0.44	0.37	0.45
Minimax-H3	0.61	0.73	0.52	0.55	0.61
HunyuanVideo	0.33	0.39	0.26	0.30	0.33
OmniWeaving	0.48	0.56	0.43	0.39	0.48
Wan2.2-TI2V-5B	0.55	0.58	0.53	0.50	0.54
Cosmos-Predict2.5	0.30	0.33	0.41	0.39	0.35
LTX-Video-2B	0.51	0.58	0.48	0.45	0.51
Video-CoCo	0.51	0.58	0.48	0.45	0.51
Video-Coder-9B + OmniWeaving	0.52	0.58	0.52	0.44	0.51
Video-Coder-9B + Minimax-H3	0.68	0.79	0.60	0.61	0.67

Table 2: Performance of different models on the VBench-2.0 benchmark. Best performance is in **bold**.
Method	Mechanics (↑)	Thermotics (↑)	Material (↑)	Average (↑)
Sora	62.22%	43.36%	64.94%	56.84%
Kling 1.6	65.55%	59.46%	68.00%	64.34%
HunyuanVideo	76.09%	56.52%	64.37%	65.66%
Minimax-H3	67.59%	47.22%	48.94%	54.58%
CogVideoX-1.5	80.80%	67.13%	83.19%	77.04%
Video-Coder-9B + Minimax-H3	88.67%	70.88%	75.02%	78.19%

Table 3: Ablation study of Video-Coder on PhyGenBench.
Methods	PhyGenBench
Minimax-H3	0.61
+ Video-Coder-SFT	0.63
+ Video-Coder-wo-annealing	0.65
+ Video-Coder-wo-video reward	0.64
+ Video-Coder-wo-execution reward	0.63
+ Video-Coder	0.67

Qualitative Examples

Prompt← Draft video (Video-Coder)Generated video (MiniMax-H3) →

Figure 3: Drag each video to compare the draft from Video-Coder's Blender code with the final video from MiniMax-H3.