Quality Tool
The Quality Tool helps you measure and review the quality of the images in a dataset, so you can keep the strongest images and cull the weak ones before training.
It has two complementary sides. The first scores images automatically, either with a fast heuristic or with a GPU quality model, and can move low-scoring images out of the dataset in a single batch. The second lets you review images by hand, assigning your own scores through ordinary rating, head-to-head comparisons, or a quick keep-or-trash pass.
Both sides write quality scores that follow the image around Studio. Once an image has been scored, you can sort by that score in Gallery View or use it as a filter when building a focused dataset with Carve.
When scoring images automatically, you can choose between two modes.
Simple Mode runs on the CPU and estimates quality from a handful of image measurements. It looks at sharpness, contrast, exposure, and the amount of detail retained relative to file size, then combines these into a single score from 0 to 100.
Because it needs no model and no GPU, Simple Mode is fast and works on any machine. It is well suited to catching obvious problems such as blurred, under-exposed, over-exposed, or heavily compressed images.
AI Mode uses a no-reference image-quality model to score each image on the same 0 to 100 scale. It runs on your GPU and generally produces judgements that track human perception more closely than the Simple heuristic, at the cost of requiring a compatible GPU and the AI dependencies.
Note: If the AI quality stack is not available, Studio falls back to Simple Mode automatically rather than failing the scan.
You can set the Batch Size used by AI Mode. Larger batches score images faster but use more VRAM, so reduce the batch size if you encounter out-of-memory errors.
AI Mode uses the built-in MANIQA model by default.
If custom quality models are present, they appear in the model list alongside the built-in model and are marked as Custom. Studio detects them automatically, so a custom model becomes available for selection as soon as it is in place.
Note: Scores are cached in the database as they are calculated. When you re-scan a folder, images that have already been scored in the selected mode are reused from the cache, so only new or unscored images need to be processed.
The scoring side of the tool follows a simple sequence.
- Choose Simple or AI mode, and select a model and batch size if you are using AI Mode.
- Select Start Scan. Studio first counts the eligible images and asks you to confirm before the scan begins.
- The scan runs in the background. Progress is shown while images are analysed, and you can continue using other parts of Studio.
- When the scan finishes, the results appear as a grid sorted from highest to lowest score. Each thumbnail carries a score badge, coloured to show quality at a glance.
The scan respects any active search filter, so you can limit scoring to a specific selection of images rather than the whole folder.
Once results exist, the sidebar controls decide which images are kept and which are marked for removal. Adjusting them updates the preview immediately without touching any files.
| Control | What it does |
|---|---|
| Min Score | Marks images below the chosen score (0–100) for removal. |
| Min Megapixels | Marks images below the chosen resolution for removal. |
| Folder Limit | Keeps only the best N images per folder. 0 disables the limit. |
| Sequence Limit | Keeps only the best N images per detected sequence, such as burst shots sharing a filename pattern. 0 disables it. |
| Validation Split | Holds back a percentage of the kept images as a validation set. |
Note: The Validation Split control only appears when the train/validation split is enabled in the dataset's settings. When it is off, no images are held back.
The Preview panel shows how many images will be kept and removed under the current settings, along with a sample of a kept and a removed image so you can sanity-check your thresholds.
When you are satisfied with the preview, Run Batch Process applies the current thresholds. Images marked for removal are moved to the dataset's .trash folder, and, if a validation split is configured, the held-out images are moved into the validation set.
This step moves files, so Studio asks you to confirm before it begins and shows its progress as it works. Removed images are moved to .trash rather than deleted outright, so the operation can be recovered from if needed.
Alongside automatic scoring, the Quality Tool lets you review images and assign your own scores by hand. Manual scores are stored separately from the AI scores but use the same 0 to 100 scale.
Where an image already has an AI quality score, that score is offered as the starting point for your rating. If you leave it unchanged, the image is passed over without overwriting anything, so you only record a score where you have actually made a judgement.
Review sessions can run in several modes.
- Sequential: Works through the images in order, which is useful for systematic review.
- Random: Shuffles the order to reduce bias.
- Versus: Compares two images head-to-head. See below.
-
Swipe: A quick keep-or-trash pass, where you swipe or use the arrow keys to keep an image or send it to
.trash.
While reviewing, you can also edit an image's caption and tags, move a mis-sorted image to the Unsorted folder for later, or delete it. Any action can be undone.
Versus mode presents two images at a time and asks you to pick the better one. This is often easier and more consistent than assigning an absolute score to each image on its own.
Behind the scenes, each image carries a rating that is adjusted after every comparison using an ELO-style calculation, similar to the ranking systems used for competitive play. The winner's score rises and the loser's falls, with larger changes when a lower-rated image beats a higher-rated one. Over many comparisons, the strongest images rise to the top.
- The winning image stays on to face a new opponent, while the loser leaves the round with its adjusted score.
- Opponents are matched by similar score, with a configurable Match Range so you can decide how close in quality the two images should be. A tighter range produces closer contests; a wider range mixes in more varied opponents.
- Studio also tries to pair images of the same orientation and, where possible, from the same folder, so the comparisons feel fair.
- A super pick (holding Shift while choosing) sends the losing image straight to
.trash, which is handy for quickly discarding a clear loser. - After a run of consecutive wins, the reigning image retires and a fresh pair begins, so a single strong image does not dominate the whole session.
- Previously trashed images are periodically reintroduced, giving rejected images a chance to earn their place back.
By default, Versus mode draws from images that have not yet been given a manual score, but you can choose to include already-rated images as well.
Quality scores are stored with each image and shared across Studio, so the effort you spend scoring and reviewing carries over into the rest of your workflow.
- In Gallery View, the Quality Score sort orders images by their score, whether that score came from automatic scoring or from manual review.
- In Carve, you can build a new dataset from images that meet a chosen score, making it straightforward to extract only your best images from a larger collection.