Human Preference Annotations for Image-to-Video Generation

3,000 rows of human preference data for evaluating image-to-video generation. Each row contains a reference image, two generated videos from Pika and CogVideoX models, and 10 aggregated human annotations. The dataset was created by datapointai and last updated in March 2026.

Use Cases

• Training reward models for RLHF by learning from the 10 human preference annotations per row.
• Benchmarking video generation models like Pika and CogVideoX using the paired video outputs and aggregated majority votes.
• Fine-tuning image-to-video models via DPO using the reference image and the preferred video output.
• Analyzing human aesthetic preferences in video generation by studying the annotation patterns across 3,000 samples.

Strengths

• Contains 3,000 complete data rows for model training and evaluation.
• Each sample includes 10 human annotations aggregated via majority vote for reliable preference labels.
• Provides direct comparisons between outputs from two prominent models, Pika and CogVideoX.

Limitations

• The dataset size of 3,000 rows may be insufficient for training large-scale models without augmentation.
• Preference annotations are aggregated, losing granularity from individual human raters.
• Geographic and demographic bias is likely, as the description notes a 'Region:us' tag.

Provenance

Source

datapointai via Hugging Face.

Collection Method

Human annotations collected for pairs of videos generated from reference images by Pika and CogVideoX models.

Freshness

Last updated March 2026.

Geography

Primarily United States, based on platform tags.