VTR-Bench

Benchmarking Text Fidelity and Instruction Following in Video Generation

Code and Annotations   |   Generated Videos   |   Video Directory Guide

Overview

Can a video generator render the requested text while preserving the scene, layout, and motion described in a prompt? VTR-Bench evaluates these two complementary aspects of video generation: text fidelity and instruction following.

The benchmark places text within advertisements, scientific scenes, user interfaces, cultural settings, and daily life. Each prompt specifies the exact text to render and its carrier, together with the surrounding visual content. Evaluation combines text transcription with a scene-specific checklist.

This Hugging Face repository hosts the generated videos. Benchmark annotations, evaluation code, and the Agentic I2V framework are available in the GitHub repository.

Benchmark Data

The benchmark contains 300 prompts, 1,202 required-text blocks, and 6,000 checklist questions, with 60 prompts per domain and 20 questions per prompt.

Domain	ID prefix	Prompts
Advertisement	AD	60
Science	SCI	60
User Interface	UI	60
Culture	CULT	60
Daily Life	LIFE	60

The final annotations are bundled with the code:

• prompts.json: original English prompts, case IDs, scene metadata, and verbatim text references.
• checklists.json: questions covering Entity Presence, Spatial Relationship, Temporal Consistency, Motion Adherence, and Scene Attributes.

Generated Videos

The collection contains 5,698 videos across 19 model or experiment folders. Each folder is directly under generated_videos/, and filenames match benchmark IDs, for example generated_videos/hunyuanvideo-1.5/AD-0001.mp4.

The video directory guide lists all folders and their generation settings. In particular:

• minimax-h3_agentic/: videos selected by the Agentic I2V framework. Every video-generation request preserves the original prompt and adds a motion refinement; the agent can inspect and revise image or video candidates.
• minimax-h3_i2v/: the first-image ablation. The earliest generated image from the corresponding Agentic run and the verbatim original prompt are passed directly to H3, without visual feedback, image editing, candidate selection, or an added motion refinement.
• The ltx_* and minimax-h3_* resolution-duration folders provide the available low/high-resolution and short/long-video settings.

All folders contain 300 videos except viduq3-pro/, which contains 298; SCI-0023 and UI-0008 are absent. Evaluate the available IDs without treating missing files as generation failures.

Download

Install the Hugging Face client, then download only the folder you need. If the repository requires authentication, first sign in with an account that has access using hf auth login.

python -m pip install -U huggingface_hub
hf download hardenyu/VTR-Bench --repo-type dataset \
  --include "generated_videos/minimax-h3_agentic/*.mp4" \
  --local-dir ./VTR-Bench-data

Replace minimax-h3_agentic with another folder from the directory guide. The download retains the repository layout. See the official Hugging Face download guide for authentication, revision selection, and download options.

Evaluation

Set up the evaluation environment and local evaluator checkpoint using the code README. Then provide the downloaded model folder as the only required input:

vtr-bench evaluate ./VTR-Bench-data/generated_videos/minimax-h3_agentic

The command runs both metrics with the bundled annotations:

• Checklist score ↑ measures instruction following through question-weighted yes rates, overall and across the five checklist dimensions.
• Word Error Rate (WER) ↓ measures transcription errors against the required text, with R+1, R+5, and R+10 hypothesis-token caps and equal weighting across videos. The evaluator does not receive the target strings while transcribing the video.

Scores are saved to results/<input-name>/metrics.json. For Agentic generation, refer to the generation instructions.

License

The repository uses the Apache-2.0 license. The underlying models retain their respective licenses and usage terms.