nori-captioner

Local Vision Caption Studio for Images and Video.

A web UI for captioning image/video datasets in-place using local VLMs. Captions are saved as sidecar .txt files next to each media file.

Quick start

uvx nori-captioner

Scans the current directory (or a given path) recursively for images and videos and opens a local web UI.

uvx nori-captioner /path/to/dataset

Features

• Recursive directory scan — hidden directories are excluded
• Per-file metadata display: resolution, duration, frame count, fps
• Manual caption editing with autosave and save-state feedback
• Upload images/videos via file picker or drag-and-drop
• Delete files (removes media and sidecar caption together)
• Caption lock support with persistent locked state per media item
• Auto-captioning queue with single-file and batch modes
• Configurable user prompt — editable in the UI and persisted to disk
• Pagination and filters for all / captioned / uncaptioned / unlocked / locked / queued

Auto-captioning with local VLMs

Run with VLM extras:

uvx --from 'nori-captioner[vlm]' nori-captioner

Note: Qwen3-VL requires torchvision, which is included in the vlm extra. On Linux x86_64, CUDA 12.8 wheels for torch and torchvision are used automatically.

Optional 4-bit / 8-bit quantization:

uvx --from 'nori-captioner[vlm,quantize]' nori-captioner

Run with a built-in model alias:

uvx --from 'nori-captioner[vlm]' nori-captioner --model qwen3-vl:8b

Or pass any Hugging Face model ID directly:

uvx --from 'nori-captioner[vlm]' nori-captioner --model your-org/your-vlm

Model aliases

CLI options

Prompt persistence

The user prompt edited in the web UI is saved to .nori-captioner.settings.json in the scanned directory and automatically restored on next launch.