Civitai As Characters: AI-Generated Character Portraits with Face Bounding Boxes

A deduplicated collection of AI-generated portrait images from CivitAI, each described as a character by the Llava1.6-34b model. Each image is taller than it is wide and contains exactly one face, with bounding box coordinates provided. The dataset was created by kubernetes-bad and last updated on February 27, 2024.

Use Cases

• Train character generation models based on the 'character-like' textual descriptions.
• Develop or benchmark face detection algorithms using the provided bounding boxes.
• Create synthetic datasets for portrait-focused computer vision tasks based on the curated portrait criteria.
• Study the alignment between AI-generated images and AI-generated textual descriptions.

Strengths

• Images are filtered to be portraits, meaning they are taller than they are wide.
• Each image contains exactly one face, with bounding boxes provided for consistency.
• Each image includes a character description generated by the Llava1.6-34b model.

Limitations

• Dataset size, row count, and file formats are unknown, limiting suitability assessment.
• Column-level documentation is absent; field semantics must be inferred after download.
• Data may reflect the stylistic and thematic biases inherent to the source platform, CivitAI.

Provenance

Source

CivitAI, a platform for sharing AI-generated images.

Collection Method

Images were searched by SD XL-derived models and described by the Llava1.6-34b model.

Freshness

Last updated 2024-02-27 22:14:17.