license: cc-by-4.0
configs:
- config_name: thumbnails
data_files: train/thumbnails/thumbnails_metadata.csv
- config_name: videos
data_files: train/videos/video_metadata.csv
- config_name: timelapses
data_files: train/timelapses/timelapses_metadata.csv
size_categories:
- 1K<n<10K
Bambu Timelapse Dataset
Dataset Summary
Disclaimer: This dataset is an independent community project and is not affiliated with or endorsed by Bambu Lab in any official capacity. We simply curate footage captured on its consumer printers to enable open research.
The Bambu Timelapse Dataset is an open, community‑driven collection of time‑lapse videos captured on Bambu Lab 3‑D printers (P1 series, X1 series and variants). Its goal is to provide a high‑quality video corpus that researchers and hobbyists can use to:
- develop print‑failure‑detection or quality‑inspection models,
- benchmark video‑classification / video‑anomaly‑detection methods,
- experiment with self‑supervised or contrastive learning on real‑world maker footage.
Why this dataset matters
- First fully open corpus – to our knowledge there are no openly‑licensed datasets that share the complete set of 3‑D‑printer timelapses used in prior failure‑detection papers or commercial products. By releasing the raw videos and metadata, we remove a major reproducibility barrier.
- Built on the market‑leading platform – Bambu Lab printers are among the most popular FDM machines today and their in‑printer timelapse feature produces millions of clips each month. Leveraging this ubiquitous hardware means the data is easy to grow and highly relevant for both researchers and hobbyists.
- Manufacturing & QA playground – beyond academic CV tasks, the dataset is a lightweight entry point for exploring real‑world manufacturing quality‑assurance pipelines: anomaly detection, digital twins, predictive maintenance and more.
At the time of writing, the dataset contains:
| Modality | Files | Raw size* | Notes |
|---|---|---|---|
Timelapse videos (video) |
627 | ~3 GB | .mp4 / .avi, various lengths |
| Full‑length camera videos | 81 | ~22 GB | direct recordings from the printer’s onboard camera system, kindly donated by a community member who archived the raw feed |
| Thumbnails / preview frames | 627 | ~40 MB | single JPG extracted per timelapse |
| CSV metadata | 3 | ≈63 kB | video_metadata.csv, timelapses_metadata.csv, thumbnails_metadata.csv |
Raw size is approximate and will grow as new contributions land. Each video is accompanied by duration, frame count, file extension and a contributor ID. No print‑outcome labels are provided yet; those will be added in a future, annotated release.
Supported Tasks & Benchmarks
- Video Classification – train a network to predict failure vs. success once labels are available.
- Frame‑level/Clip‑level Anomaly Detection – detect shifts in nozzle path, spaghetti, layer shifts, etc.
- Self‑supervised Representation Learning – learn robust 3‑D‑printer visual embeddings without annotations.
- Visual Build‑Plate Identification – automatically recognize plate types (textured PEI, smooth PEI, glass, etc.) to improve domain adaptation across different printer setups.
No public leaderboard exists yet; feel free to open a PR to add one!
Roadmap & Planned Models
The first reference model scheduled for release will tackle build‑plate identification/detection end‑to‑end. It is intended as a lightweight, memory‑efficient network that can run directly on the P1S/X1C’s embedded environment, providing a visual "sanity check" that the plate type Bambu Studio expects actually matches what is sitting on the printer before a job is sent.
Dataset Structure
bambu-timelapse-dataset/
│
├── metadata/
│ ├── timelapses_metadata.csv
│ ├── video_metadata.csv
│ └── thumbnails_metadata.csv
│
└── originals/
├── timelapses/<contributor_id>/*.mp4
├── videos/<contributor_id>/*.mp4
└── thumbnails/<contributor_id>/*.jpg
Key fields
| Field | Type | Description |
|---|---|---|
video |
datasets.features.Video() |
Timelapse clip (variable FPS & length) |
contributor |
— | Identifier for the user who provided the timelapse. Used to group clips by origin and track contributions. |
duration |
float |
Length in minutes (metadata CSV) |
num_frames |
int |
Number of frames in the file |
extension |
string |
.mp4 or .avi |
Splits
| Split | # rows |
|---|---|
train |
627 |
Usage
from datasets import Video, load_dataset
from decord.video_reader import VideoReader
import matplotlib.pyplot as plt
# Load the dataset (timelapse split)
data = load_dataset("v2thegreat/bambu-timelapse-dataset", "timelapses")
# Decode the 101st video (index 100) from the "file_name" column
frame: VideoReader = data["train"].cast_column("file_name", Video())[100]["file_name"]
# Display the final frame of the video
plt.imshow(frame[-1].asnumpy())
plt.show()
How to Contribute
- Download your Bambu timelapse from Bambu Studio, Handy App, or via FTP.
- Check that it only contains the build plate (crop or blur personal items if needed).
- Reach out to @v2thegreat via the Hugging Face "Community" tab or, even easier, open a Pull Request against this repo that adds your timelapses under
originals/timelapses/<your_id>/. - (Optional) Include a CSV row with the columns
user,duration,extension,num_frames,relative_path.
Future releases will provide a small, labeled subset so the community can bootstrap better detectors.
Licensing
All content is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to use the dataset in commercial or academic projects provided you give appropriate credit.
Citation
@misc{bambu_timelapse_2025,
title = {Bambu Timelapse Dataset},
author = {Abbas, Ali and community contributors},
year = {2025},
howpublished = {\url{https://huggingface.co/datasets/v2thegreat/bambu-timelapse-dataset}},
note = {Version 1.0},
license = {CC-BY-4.0}
}
Acknowledgements
This project started with a Reddit call‑for‑data – thank you to everyone who donated timelapses and ideas.
Last updated: 2025‑04‑17 – 627 timelapses from 3 contributors.