💡 Feature Requests and Ideas for LightlyStudio

[IgorSusmelj]

We would love your feedback on what to build next for LightlyStudio.

🗳️ How to participate:

• Drop your feature ideas or requests as a comment below.
• Upvote the comments that match what you want to see in LightlyStudio first.

We will regularly review the most upvoted ideas and keep you updated as we roll out new features.

Your feedback helps shape the future of open-source data tooling. Thank you for being part of the journey!

[faileon]

Dockerfile would be nice.

[bw4sz]

A gallery-view for object detection would be huge for my community. Imagine you have a model that makes object detections and you want to classify them. Most annotation tools ask you to do this on a per-image basis, but in tasks where objects are small and you have thousands of images, there is a ton of zooming and panning. A module that groups similar objects by label and image features, allows you select a bunch of them and label them could be huge in rapid sorting. You could always follow up on a per-image basis. There is something like this in the paid versus of label-studio, but not in the free version, a sure fire sign its critical to many users.

Same idea, but could present this in embedding space, you could select and a bunch of points and relabel them. "Hey this cluster is actually 'bear'". Tools that facilitate rapid relabeling that is then followed by fine-grained labeling.

Mocked up something here

I am the developer of an open source machine learning library for airborne biodiversity object detection.

https://deepforest.readthedocs.io/

Images like

or dense images of vegetation -> https://visualize.idtrees.org/

I am thinking about more tightly integrating an active learning module and annotation interface. We currently use label-studio.

Another feature request is double click on label or annotation and zooms to box. Label-studio said no to this (can find the issue if interested).

[eternal-f1ame]

Add automatic folder-based tagging to add_samples_from_path()

Summary

Add tag_depth parameter to add_samples_from_path() to automatically tag images based on their folder hierarchy, enabling easy organization of datasets with nested folder structures.

Motivation

Many datasets are organized hierarchically by class, subtype, or other categories:

data/
├── dogs/
│   ├── golden_retriever/
│   └── husky/
├── cats/
│   ├── persian/
│   └── siamese/

Currently, users must manually tag images after loading, which is tedious and error-prone. Automatic folder-based tagging would make dataset organization seamless.

Proposed Solution

Add optional tag_depth parameter (default: 0) to add_samples_from_path():

dataset.add_samples_from_path(
    path="data/",
    tag_depth=2  # Tag based on first 2 folder levels
)

Behavior:

• tag_depth=0 (default): No automatic tagging, maintains backward compatibility
• tag_depth=1: Tags by immediate parent folder only
• tag_depth=2: Tags by first two folder levels (multiple tags per image)
• tag_depth=N: Tags by first N folder levels

Example with tag_depth=2:

data/
├── dogs/
│   ├── golden_retriever/
│   │   └── img1.jpg  → tags: ["dogs", "golden_retriever"]
│   └── husky/
│       └── img2.jpg  → tags: ["dogs", "husky"]
├── cats/
│   └── img3.jpg      → tags: ["cats"]
└── img4.jpg          → tags: []

API Changes

def add_samples_from_path(
    self,
    path: PathLike,
    allowed_extensions: Iterable[str] | None = None,
    embed: bool = True,
    tag_depth: int = 0,  # NEW PARAMETER
) -> None:

Implementation Details

• Automatically creates tags for each folder level up to tag_depth
• Each tag is created once and shared across all applicable images
• Images can have multiple tags (one per folder level)
• Images directly in base path are not tagged
• Fully backward compatible (default tag_depth=0)

Benefits

• Zero manual tagging effort for hierarchical datasets
• Supports complex taxonomies (class → subclass → variant)
• Works with existing YOLO/COCO workflows
• Enables easy filtering and organization in GUI

Use Cases

• Classification datasets organized by class/subclass
• Medical imaging with patient/study/series structure
• Multi-label datasets with hierarchical categories
• Splitting work by category for annotation teams

[masakljun]

Hi @eternal-f1ame

We've just merged a PR that implements the first part of your feature request for directory-based tagging.

The add_samples_from_path() function now includes a tag_depth=1 parameter, which automatically creates tags from the parent directory names.

You can test it by installing the library from the main branch:
pip install git+https://github.com/lightly-ai/lightly-studio.git
If you're interested, you could also look into extending this to support tag_depth > 1 (for example, tagging from grandparent directories) or apply the same logic to the add_samples_from_yolo and add_samples_from_coco functions.

Thanks again for suggesting this improvement 😄

[eternal-f1ame]

@masakljun lovely, I will check this out over the weekend

[alpha027]

Currently the app uses model to perform data selection, does it support model-guided labeling ( using open-source or custom models) ?
For instance use the Segment Anything (SAM) model to help label the raw or selected data, or some classification model to automatically generate labels for the dataset