Deepbooru-Tagwalker Walks Tags First to Simplify Dataset Verification

By vramkickedin | | 2 min read
A large frosted glass tag icon engraved with a subtle checkmark floats above a curved trail of tiny square image frames.

Deepbooru-tagwalker is a lightweight desktop tool that improves the accuracy of existing tags in image datasets. Instead of opening images and editing tags one by one, you select a single tag and the program walks you through every image in sequence, asking a simple yes or no question. It handles large collections without slowdown and requires no Python installation or technical setup.

Developer Elliezrah built the tool as a personal solution to a problem they encountered during years of dataset tagging work. Finding no existing software that approached tag verification this way, they created a tag-centric workflow that checks consistency across an entire dataset. The project holds a special place as their first experiment with prompt-driven development, using Qwen and Claude models to generate every line of code.

First major release brings tag-first proofing to dataset preparation

Key features
  • Sequential tag-first review of entire datasets.
  • Color-coded image queue with undo support.
  • Instant text file updates on each decision.
  • Clickable queue for jumping to any image.
  • Zoom popup for detailed image inspection.
  • Skips and handles uncertain images gracefully.
  • Runs as standalone executable, no dependencies.
  • Works with standard Stable Diffusion dataset formats.

This tool fits people training LoRA models or fine-tuning checkpoints who need verified tag accuracy, not just tag presence. It serves anyone auditing large batches of images tagged by automated tools, where consistency determines training quality. The workflow eliminates the common problem of losing track of which images you have already checked across long sessions.

Developer notes and current limitations

The release comes with some important gaps that Elliezrah plans to address in future versions. There is no tag search or filtering, so all tags load directly into the sidebar without organization, and the program does not save progress between sessions. Accidentally clicking a different tag mid-review will immediately switch context without warning, causing you to lose your current position.

"This is my first ever vibe coding project. I have zero programming knowledge. The initial version was built entirely through prompting — using Qwen 3.6 Q4 for the first draft, then Claude Sonnet 4.6 for refinement, debugging, and performance work." — Source: GitHub

Project Links