text-2-video-human-preferences-luma-ray2

<style>

.vertical-container {
display: flex;
flex-direction: column;
gap: 60px;
}

.image-container img {
height: 150px; /* Set the desired height */
margin:0;
object-fit: contain; /* Ensures the aspect ratio is maintained */
width: auto; /* Adjust width automatically based on height */
}

.image-container {
display: flex; /* Aligns images side by side */
justify-content: space-around; /* Space them evenly */
align-items: center; /* Align them vertically */
}

.container {
width: 90%;
margin: 0 auto;
}

.text-center {
text-align: center;
}

.score-amount {
margin: 0;
margin-top: 10px;
}

.score-percentage {
font-size: 12px;
font-weight: semi-bold;
}

</style>

# Rapidata Video Generation Luma Ray2 Human Preference

<a href="https://www.rapidata.ai">
<img src="https://cdn-uploads.huggingface.co/production/uploads/66f5624c42b853e73e0738eb/jfxR79bOztqaC6_yNNnGU.jpeg" width="300" alt="Dataset visualization">
</a>

<a href="https://huggingface.co/datasets/Rapidata/text-2-image-Rich-Human-Feedback">
</a>

<p>
If you get value from this dataset and would like to see more in the future, please consider liking it.
</p>

This dataset was collected in ~1 hour total using the [Rapidata Python API](https://docs.rapidata.ai), accessible to anyone and ideal for large scale data annotation.

# Overview

In this dataset, ~45'000 human annotations were collected to evaluate Luma's Ray 2 video generation model on our benchmark. The up to date benchmark can be viewed on our [website](https://www.rapidata.ai/leaderboard/video-models).
The benchmark data is accessible on [huggingface](https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences) directly.

# Explanation of the colums

The dataset contains paired video comparisons. Each entry includes 'video1' and 'video2' fields, which contain links to downscaled GIFs for easy viewing. The full-resolution videos can be found [here](https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences/tree/main/Videos).

The weighted_results column contains scores ranging from 0 to 1, representing aggregated user responses. Individual user responses can be found in the detailedResults column.

# Alignment

The alignment score quantifies how well an video matches its prompt. Users were asked: "Which video fits the description better?".

## Examples

<div class="vertical-container">
<div class="container">
<div class="text-center">
<q>A lone kayaker paddles through calm, reflecting waters under a vibrant sunset, the sky painted with hues of orange and pink, creating a serene and mesmerizing evening scene.</q>
</div>
<div class="image-container">
<div>
<h3 class="score-amount">Ray 2 </h3>
<div class="score-percentage">(Score: 91.56%)</div>
<img src="https://assets.rapidata.ai/0046_ray2_1.gif" width=500>
</div>
<div>
<h3 class="score-amount">Hunyuan </h3>
<div class="score-percentage">(Score: 8.44%)</div>
<img src="https://assets.rapidata.ai/0046_hunyuan_1724.gif" width=500>
</div>
</div>
</div>
<div class="container">
<div class="text-center">
<q>A sunset view over a bustling Tokyo street, neon lights flickering as crowds weave through the vibrant night. Capture reflections on wet pavement and the dynamic energy of city life as day transitions to night.</q>
</div>
<div class="image-container">
<div>
<h3 class="score-amount">Ray 2 </h3>
<div class="score-percentage">(Score: 2.83%)</div>
<img src="https://assets.rapidata.ai/0063_ray2_1.gif" width=500>
</div>
<div>
<h3 class="score-amount">Sora </h3>
<div class="score-percentage">(Score: 97.17%)</div>
<img src="https://assets.rapidata.ai/0063_sora_1.gif" width=500>
</div>
</div>
</div>
</div>

# Coherence

The coherence score measures whether the generated video is logically consistent and free from artifacts or visual glitches. Without seeing the original prompt, users were asked: "Which video is logically more coherent? E.g. the video where physics are less violated and the composition makes more sense."

## Examples

<div class="vertical-container">
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">(Score: 90.42%)</div>
<img src="https://assets.rapidata.ai/0098_ray2_1.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Pika </h3>
<div class="score-percentage">(Score: 9.58%)</div>
<img src="https://assets.rapidata.ai/0098_pika_2445694862.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">(Score: 4.11%)</div>
<img src="https://assets.rapidata.ai/0086_ray2_2.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Pika </h3>
<div class="score-percentage">(Score: 95.89%)</div>
<img src="https://assets.rapidata.ai/0086_pika_1678426151.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
</div>

# Preference

The preference score reflects how visually appealing participants found each video, independent of the prompt. Users were asked: "Which video do you prefer aesthetically?"

## Examples

<div class="vertical-container">
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">(Score: 61.12%)</div>
<img src="https://assets.rapidata.ai/0036_ray2_1.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Sora </h3>
<div class="score-percentage">(Score: 38.88%)</div>
<img src="https://assets.rapidata.ai/0036_sora_1.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">(Score: 39.48%)</div>
<img src="https://assets.rapidata.ai/0020_ray2_2.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Hunyuan </h3>
<div class="score-percentage">(Score: 60.52%)</div>
<img src="https://assets.rapidata.ai/0020_hunyuan_1724.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
</div>

</br>

# About Rapidata

Rapidata's technology makes collecting human feedback at scale faster and more accessible than ever before. Visit [rapidata.ai](https://www.rapidata.ai/) to learn more about how we're revolutionizing human feedback collection for AI development.

# Other Datasets

We run a benchmark of the major image generation models, the results can be found on our [website](https://www.rapidata.ai/leaderboard/image-models). We rank the models according to their coherence/plausiblity, their aligment with the given prompt and style prefernce. The underlying 2M+ annotations can be found here:

- Link to the [Rich Video Annotation dataset](https://huggingface.co/datasets/Rapidata/text-2-video-Rich-Human-Feedback)
- Link to the [Coherence dataset](https://huggingface.co/datasets/Rapidata/Flux_SD3_MJ_Dalle_Human_Coherence_Dataset)
- Link to the [Text-2-Image Alignment dataset](https://huggingface.co/datasets/Rapidata/Flux_SD3_MJ_Dalle_Human_Alignment_Dataset)
- Link to the [Preference dataset](https://huggingface.co/datasets/Rapidata/700k_Human_Preference_Dataset_FLUX_SD3_MJ_DALLE3)

We have also colleted a [rich human feedback dataset](https://huggingface.co/datasets/Rapidata/text-2-image-Rich-Human-Feedback), where we annotated an alignment score of each word in a prompt, scored coherence, overall aligment and style preferences and finally annotated heatmaps of areas of interest for those images with low scores.

<style>
.vertical-container {
display: flex;
flex-direction: column;
gap: 60px;
}

.image-container img {
height: 150px; /* Set the desired height */
margin:0;
object-fit: contain; /* Ensures the aspect ratio is maintained */
width: auto; /* Adjust width automatically based on height */
}

.image-container {
display: flex; /* Aligns images side by side */
justify-content: space-around; /* Space them evenly */
align-items: center; /* Align them vertically */
}

.container {
width: 90%;
margin: 0 auto;
}

.text-center {
text-align: center;
}

.score-amount {
margin: 0;
margin-top: 10px;
}

.score-percentage {
font-size: 12px;
font-weight: semi-bold;
}
</style>

# Rapidata 视频生成 Luma Ray2 人类偏好数据集

<a href="https://www.rapidata.ai">
<img src="https://cdn-uploads.huggingface.co/production/uploads/66f5624c42b853e73e0738eb/jfxR79bOztqaC6_yNNnGU.jpeg" width="300" alt="Dataset visualization">
</a>

<a href="https://huggingface.co/datasets/Rapidata/text-2-image-Rich-Human-Feedback">
</a>

<p>若您从本数据集获益并希望后续获取更多同类资源，不妨为其点赞。</p>

本数据集总计耗时约1小时，通过[Rapidata Python API](https://docs.rapidata.ai)完成采集，该工具面向所有用户开放，非常适合大规模数据标注工作。

# 数据集概览

本数据集共采集约45000条人类标注结果，用于在我们的基准测试中评估Luma的Ray2视频生成模型。最新版基准测试可通过我们的[官网](https://www.rapidata.ai/leaderboard/video-models)查看，基准测试数据可直接在[Hugging Face](https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences)获取。

# 字段说明

本数据集包含成对的视频对比样本。每条数据均包含`video1`与`video2`字段，其中存储了用于快速预览的压缩GIF文件链接。全分辨率视频可通过[此链接](https://huggingface.co/datasets/Rapidata/text-2-video-human-preferences/tree/main/Videos)获取。`weighted_results`字段包含0至1区间的分数，代表聚合后的用户反馈结果；单条用户标注结果则可在`detailedResults`字段中查看。

# 对齐度

对齐度分数用于量化视频与提示词的匹配程度。调研向参与者提出的问题为：「哪一段视频更贴合给定的描述？」

## 示例

<div class="vertical-container">
<div class="container">
<div class="text-center">
<q>「一名独自划桨的皮划艇运动员在平静如镜的水面穿行，绚烂日落之下，天空被橙粉交织的色彩晕染，营造出静谧迷人的黄昏景致。」</q>
</div>
<div class="image-container">
<div>
<h3 class="score-amount">Ray 2 </h3>
<div class="score-percentage">（得分：91.56%）</div>
<img src="https://assets.rapidata.ai/0046_ray2_1.gif" width=500>
</div>
<div>
<h3 class="score-amount">Hunyuan </h3>
<div class="score-percentage">（得分：8.44%）</div>
<img src="https://assets.rapidata.ai/0046_hunyuan_1724.gif" width=500>
</div>
</div>
</div>
<div class="container">
<div class="text-center">
<q>「东京繁华街道的日落景致，霓虹灯光在热闹的夜色中闪烁，人群穿梭其中。捕捉雨后湿润路面的倒影，以及昼夜交替时城市生活的蓬勃活力。」</q>
</div>
<div class="image-container">
<div>
<h3 class="score-amount">Ray 2 </h3>
<div class="score-percentage">（得分：2.83%）</div>
<img src="https://assets.rapidata.ai/0063_ray2_1.gif" width=500>
</div>
<div>
<h3 class="score-amount">Sora </h3>
<div class="score-percentage">（得分：97.17%）</div>
<img src="https://assets.rapidata.ai/0063_sora_1.gif" width=500>
</div>
</div>
</div>
</div>

# 连贯性

连贯性分数用于评估生成视频的逻辑自洽性，以及是否存在视觉 artifacts（伪影）或画面瑕疵。调研在不展示原始提示词的前提下，向参与者提出问题：「哪一段视频的逻辑连贯性更强？例如，物理规则违背更少、画面构图更合理的视频。」

## 示例

<div class="vertical-container">
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">（得分：90.42%）</div>
<img src="https://assets.rapidata.ai/0098_ray2_1.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Pika </h3>
<div class="score-percentage">（得分：9.58%）</div>
<img src="https://assets.rapidata.ai/0098_pika_2445694862.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">（得分：4.11%）</div>
<img src="https://assets.rapidata.ai/0086_ray2_2.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Pika </h3>
<div class="score-percentage">（得分：95.89%）</div>
<img src="https://assets.rapidata.ai/0086_pika_1678426151.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
</div>

# 审美偏好

偏好分数用于反映参与者对各段视频的视觉吸引力评价，不受提示词约束。调研向参与者提出的问题为：「从审美角度出发，你更偏好哪一段视频？」

## 示例

<div class="vertical-container">
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">（得分：61.12%）</div>
<img src="https://assets.rapidata.ai/0036_ray2_1.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Sora </h3>
<div class="score-percentage">（得分：38.88%）</div>
<img src="https://assets.rapidata.ai/0036_sora_1.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
<div class="container">
<div class="image-container">
<div>
<h3>Ray 2 </h3>
<div class="score-percentage">（得分：39.48%）</div>
<img src="https://assets.rapidata.ai/0020_ray2_2.gif" width="500" alt="Dataset visualization">
</div>
<div>
<h3>Hunyuan </h3>
<div class="score-percentage">（得分：60.52%）</div>
<img src="https://assets.rapidata.ai/0020_hunyuan_1724.gif" width="500" alt="Dataset visualization">
</div>
</div>
</div>
</div>

</br>

# 关于 Rapidata

Rapidata的技术让大规模人类反馈采集工作比以往更快速、更易获取。访问[rapidata.ai](https://www.rapidata.ai/)，了解我们如何革新AI开发中的人类反馈采集流程。

# 其他数据集

我们针对主流图像生成模型开展了基准测试，测试结果可在我们的[官网](https://www.rapidata.ai/leaderboard/image-models)查看。我们将根据模型的连贯性/合理性、与提示词的对齐度以及风格偏好对模型进行排名。相关的200万+条标注数据可通过以下链接获取：

- [Rich Video Annotation 数据集](https://huggingface.co/datasets/Rapidata/text-2-video-Rich-Human-Feedback)
- [连贯性数据集](https://huggingface.co/datasets/Rapidata/Flux_SD3_MJ_Dalle_Human_Coherence_Dataset)
- [文图对齐数据集](https://huggingface.co/datasets/Rapidata/Flux_SD3_MJ_Dalle_Human_Alignment_Dataset)
- [人类偏好数据集](https://huggingface.co/datasets/Rapidata/700k_Human_Preference_Dataset_FLUX_SD3_MJ_DALLE3)

我们还采集了[Rich Human Feedback 数据集](https://huggingface.co/datasets/Rapidata/text-2-image-Rich-Human-Feedback)，针对提示词中的每个单词标注对齐度分数，同时对图像的连贯性、整体对齐度以及风格偏好进行评分，并最终为低分图像生成感兴趣区域的热力图标注。