Analysis Tool
The Analysis Tool gives you a bird's-eye view of a dataset before you train on it. Rather than acting on individual images, it summarises the whole dataset — how many files are captioned, scored, or masked, how your images are distributed by resolution and aspect ratio, how long your captions are, and which groups and subjects appear most often. The aim is to help you spot gaps and imbalances early, so you can fix them with Studio's other tools.
Note: The Analysis Tool is an early, experimental tool and is available to Patreon subscribers. Link your account in Settings to unlock it. It is also under active development — some capabilities, such as per-dataset configuration, are still to arrive.
The Analysis Tool computes automatically the first time you open it for a dataset. Depending on the size of the dataset, the initial pass may take a little while; a progress indicator shows how many images have been analysed along with an estimated time remaining. You can carry on using other parts of Studio while it works.
Results are cached, so returning to the tool is instant. If the underlying data changes — for example after captioning, scoring, or masking a batch of images — the tool marks its results as out of date and offers a Recompute button so you can refresh them when you are ready.
The tool has two tabs, selectable from the sidebar:
- Overview — counts, distributions, and per-folder health.
- Biases — group and subject balance across the dataset.
A third tab, Configure, is shown but not yet available. It will let you set per-dataset thresholds and weights in a future release.
The Overview tab summarises the dataset as a whole.
A row of tiles at the top gives you the headline counts, each with the percentage of files it represents:
| Tile | What it counts |
|---|---|
| Files | Training-eligible files in the dataset. |
| Videos | Video files, shown for information only and excluded from the training figures. |
| Captioned | Files that carry at least one caption. |
| Quality Scored | Files that have received a quality score. |
| Captions Rated | Files that have a caption score. |
| Files Masked | Files that have a mask associated with them. |
Three bar charts show how your images and captions are spread out:
- Aspect Ratio Distribution: Groups images into the same aspect-ratio buckets OneTrainer uses at training time, so you can see which shapes dominate the dataset.
- Image Size Distribution: Shows the spread of image area in megapixels. A callout highlights how many images fall below a low-resolution threshold, which are often the first candidates for removal.
- Caption Length Distribution: Shows the number of words per caption line. Unusually long captions collapse into a final outlier bin so a handful of very long lines don't distort the chart.
Each bar is clickable. Selecting one opens the matching images in Gallery View, already filtered to that slice of the dataset and sorted worst-first where it makes sense — for example, smallest images at the top of the resolution slice. This lets you move straight from spotting a problem in the chart to reviewing the images behind it.
The Dataset Health table breaks the figures down folder by folder, so you can see which parts of the dataset are well prepared and which still need work. Each folder shows its total files, image count, how many are captioned (split into natural-language and tag captions), how many are quality- and caption-scored, and an overall health score.
The health score is a single figure that weights the presence of natural-language captions most heavily, followed by tag captions, caption scores, and quality scores. It is colour-coded so you can judge a folder's readiness at a glance. Folders expand to reveal their subfolders, and you can click any column header to sort sibling folders by that value.
The Biases tab helps you understand how balanced your dataset is across different subjects and attributes. It scans your captions for mentions of known terms and groups the results, so you can see, for example, whether one age range, gender, or category is heavily over- or under-represented.
Available groups are listed under Engine Groups in the sidebar. Selecting a group shows a Class Distribution chart and a table listing each class with its image count and share of the dataset. Each image is counted at most once per class, so the figures reflect how many images mention a class rather than how often a term appears.
Classes with only a small number of matching images are flagged with a warning, since results drawn from a small sample can be misleading. For age, a toggle lets you switch between Exact age — a per-year histogram — and Age groups, which collapses ages into broader bands.
As with the Overview charts, clicking a bar opens the matching images in Gallery View so you can review or rebalance them directly.
Note: The Biases tab relies on an analysis database to group and classify your files. If it is not available, the tab will ask you to set its path in Settings and try again.
The Analysis Tool does not create captions, scores, or masks of its own — it reports on the work you have already done with Studio's other tools. To fill in the gaps it highlights, use the tools that produce each kind of data:
- Add captions with the Captioning tools.
- Assign quality scores with the Quality Tool.
Once you have made changes, recompute the analysis to see the dataset's health improve.