jabberjabberjabber/ImageIndexer

What it solves

LLMII is designed to solve the problem of manually labeling and indexing large collections of images. Instead of manually typing captions and keywords, it uses a local visual language model to automatically generate descriptive metadata for images in a directory tree, making them searchable and organized using any standard metadata-compatible tool.

How it works

The tool runs a visual language model (VLM) locally on the user's machine via KoboldCpp. It analyzes images and generates a list of keywords and a caption for each file. This information is then written directly into the image's metadata (using ExifTool) or saved to a sidecar XMP file. To avoid reprocessing files, the tool uses specific metadata fields (Identifier and Status) to track the processing state of each image.

Who it’s for

Photographers and digital asset managers who want to automate image tagging and captioning without relying on cloud services or external databases, and who prefer to keep their data and processing entirely local.

Highlights

• Local AI Processing: Operates entirely on-device using GPU acceleration (Apple Metal, Nvidia CUDA, or AMD Vulkan) to ensure privacy and no cloud dependency.
• Metadata Integration: Writes to common metadata fields (MWG:Keyword, MWG:Description) and supports sidecar files for hashed files.
• Multi-Format Support: Compatible with a wide range of image formats, including major raw camera files.
• Highly Configurable: Offers detailed control over sampler options (temperature, top_p, top_k), image dimensions, and keyword cleaning rules (e.g., depluralization).
• Cross-Platform: Works on Windows, macOS (ARM), and Linux.