Midjourney v6 Recaptioned: 1.2M Images with Triple VLM Annotations

Comprising 1,235,432 Midjourney v6 images paired with captions generated by three different Vision Language Models (VLMs), released by Photoroom in March 2026. It provides a large-scale collection of AI-generated art with multi-perspective textual descriptions from LLaVA, Gemini Flash 1.5, and Qwen3 VL 8B. The data is formatted in Parquet for efficient processing in machine learning workflows.

Use Cases

• Training image-to-text models using the 'gemini' or 'qwen3' columns as high-quality synthetic targets
• Benchmarking VLM captioning styles by comparing the 'llava' descriptions against 'qwen3' outputs for the same image
• Fine-tuning text-to-image diffusion models to improve prompt adherence using multi-VLM consensus

Strengths

• 1,235,432 image-text pairs
• Triple-model captioning for over 1 million records
• MIT licensed for broad usability
• Uses high-quality Midjourney v6 source imagery

Limitations

• Contains 100% synthetic content which may inherit biases from both the image generator and the captioning VLMs
• Missing caption data for approximately 218,000 images in the Gemini and Qwen3 fields

Provenance

Source

Photoroom, based on the brivangl/midjourney-v6-llava dataset

Collection Method

Synthetic generation via Midjourney v6 and automated annotation using LLaVA, Gemini Flash 1.5, and Qwen3 VL 8B

Freshness

Last updated March 13, 2026.