Welcome! SEA-VL is a global community project organized by the SEACrowd community to push the boundaries of vision and language research in Southeast Asia (SEA). We recently completed Phase 1 of this project with the release of high-quality, culturally relevant image datasets and a benchmarking study for SEA languages, as detailed in this paper.
For Phase 2, we have an even bigger ambition: we want to build an open-source state-of-the-art vision language model (VLM) for Southeast Asia! We believe it’s high time to create a model that truly understands Southeast Asia culture and language, the way we see it. We want the model to reflect the visual and linguistic richness of the SEA region. This effort requires diverse contributions: high-quality data curation and annotation, prompting, model training, evaluation, and more.
Whether you’re a researcher, developer, artist, linguist, photographer, student, or just passionate about Southeast Asia, there’s a place for you in this project. Read on below!
Why Contribute?
Contributing to SEA-VL Phase 2 will give you a chance to work with an international group of researchers who are equally passionate about SEA research.
Your efforts won’t go unappreciated! Every contribution to any SEA-VL Phase 2 tasks below will earn points. Reaching 200 or more points guarantees a certificate of participation and merchandise (t-shirt and keychain), while reaching 300 or more points will earn co-authorship in the final project paper.
How can I contribute?
SEA-VL Phase 2 is split into five main tasks described below. You can contribute to any task that aligns with your interest or expertise.
Important: Read the main contributor guidelines here for more details about the tasks.
If you have any questions about the process, please join our Discord to ask questions.
Task 1: Submit a SEA Culturally-Relevant Image with Description
To build the vision-language model, we need a compilation of culturally relevant images representing all 11 Southeast Asian countries.
What qualifies as a “culturally relevant image”?
Any image that reflects an aspect of SEA culture is welcome! This could include food and cuisines (e.g., nasi goreng, pho, green curry), locations (e.g., Manila’s Escolta Street), events (e.g., Lunar New Year festivities), or everyday cultural practices (e.g., eating with your hands). As long as it connects to SEA culture, it’s a great fit!
You may check this resource to see what images we want for this task.
Go to this form to submit your images.
For bulk image upload, follow this guide!
Task 2: Review Image-Description Pairs
To ensure quality images for the dataset used in training the model, we need local annotators to rate the submitted images from Task 1.
Contributors for this task must first pass this short screening test. Check our annotation guideline here to learn more and to apply for the screening test. You will receive a link to the annotation platform if you pass the screening test.
Task 3: Translate Benchmark Datasets
To ensure the proper evaluation and testing of vision language capabilities, we need help translating prompts from existing vision language model benchmarks, such as the Aya Vision Benchmark by Cohere Labs.
Contributors for this task will translate the English prompts into any one of the following languages in which they are fluent/native speakers:
Thai, Standard Malay, Filipino/Tagalog, Lao, Khmer, Tamil, Mandarin Chinese, Burmese, Tetun, Bruneian Malay
We need three contributors per language for this task. Go to this form to contribute.
Task 4: Submit Visual Questions for Multicultural Images
To evaluate the visual understanding of the models we will train, we need to compile a dataset of high-quality questions derived from cultural images from SEA.
Contributors for this task will create original questions related to a given image (e.g., “What sport is played in this image?”) in any of the following languages in which they are fluent/native speakers:
Indonesian, Thai, Standard Malay, Filipino/Tagalog, Tamil, Chinese Mandarin, Vietnamese, Burmese, Lao, Khmer, Tetun, Bruneian Malay
Go to this form to contribute.
Task 5: Submit High-Quality Text Prompts for Image Generation
To evaluate the image generation capabilities of the models we will train, we need to compile a dataset of high-quality English-only prompts at three complexity levels.
To ensure the quality of the prompts, we need contributors who are natives or extremely familiar with the cultures of the SEA countries (Indonesia, Singapore, Philippines, Thailand, Malaysia, Vietnam, Brunei, Timor Leste, Cambodia, Laos, Myanmar)
Example prompts for the Indonesian culture:
Level 1 (Easy) - “Draw an image of people drinking cendol.”
Level 2 (Medium) - “Draw an image of people drinking cendol with durian topping.”
Level 3 (Hard) - “Draw an image of people drinking cendol with durian topping while wearing kebaya.”
Go to this form to contribute.
Task 6: Submit High-Quality Text Prompts (To be opened at a later date)
Contribution Point System for Tasks
Each task has its corresponding point-per-submission, calibrated for task difficulty and its relation to a specific SEA country culture.
We give more weight to submissions (e.g., culturally relevant images for Task 1) related to Brunei, East Timor, Cambodia, Laos, and Myanmar, as they are heavily underrepresented compared to the other countries. We also want contributors from these countries to reach the contribution threshold faster.
| Tasks | Indonesia, Singapore, Philippines | Thailand, Malaysia, Vietnam | Brunei, East Timor, Cambodia, Laos, Myanmar |
|---|---|---|---|
| Task 1: Submit a SEA Culturally-Relevant Image | 0.5 | 1 | 3 |
| Task 2: Review Image-Description Pairs | 1 | 1 | 1 |
| Task 3: Translate Benchmark Datasets | 1 | 1 | 1.5 |
| Task 4: Submit Visual Questions for Multicultural Images | 1.5 | 1.5 | 2 |
| Task 5: Submit High-Quality Text Prompts for Image Generation | 1 | 1 | 1.5 |
Remember, reaching 200 or more points will guarantee a certificate of participation and merchandise (t-shirt and keychain), while reaching 300 or more points will earn co-authorship in the final project paper.
The contribution tracking can be viewed here, where it is updated every weekend.
Project Timeline
-
May 8, 2025 - Project public launch. Contributions for Tasks 1, 2, 3, 4, and 5 are open.
-
November 8, 2025 - End of public contributions to Tasks 1, 2, 3, 4, and 5.
-
January 2026 - Prepare models, data, and multiple paper releases.
-
February 2026 - Prepare experiment paper submissions to ACL 2026. Publication of the project paper on the SEACrowd website and arXiv.
Join the Community!
Check out our GitHub page, and join our Discord server. Everyone is welcome to discuss and ask questions there!
FAQs
-
What qualifies as a “culturally-relevant image”?
- Any image that reflects an aspect of SEA culture is welcome! This could include food (e.g., eating Nasi Goreng), locations (e.g., Manila’s Escolta Street), events (e.g., Lunar New Year festivities), or everyday cultural practices (e.g., eating with your hands). As long as it connects to SEA culture, it’s a great fit!
-
Can I submit an image I posted online if I still own the copyright to it?
-
Yes, as long as you took the image and still hold the copyright.
-
All images will be openly licensed under the CC-BY-SA 4.0 license, so please ensure you own full rights to them before submission.
-
-
Do images need to be high quality?
- No, phone-quality images are perfectly acceptable as long as they’re not blurry or obstructed.
-
Can I submit images that reflect SEA culture but were taken outside of SEA?
- Yes, images taken abroad are welcome if they are culturally relevant.
-
Do I have to be a resident of the SEA culture represented in the photo I submit?
- No, you do not.
-
My native language is not a SEA indigenous language. Can I still submit my SEA culturally-relevant images?
- Yes, you can select Other… as your native language in the submission form.
-
Do you have some materials we can use to promote SEA-VL to our friends?
- Coming soon!
-
If I contribute to the open tasks but don’t participate in the modeling or experiments, will I be listed as a co-author on the paper(s) submitted to ACL 2026?
- No, contributors to the open tasks will not be listed on the ACL 2026 experimental papers. Those papers will primarily highlight model development results. However, if you reach 300 points, you will be credited as a co-author in our final organizational-wide publication, which will summarize the entire SEA-VL and other SEACrowd projects in 2025. This paper will be published on the SEACrowd website and arXiv, recognizing all community contributors who helped build the dataset and evaluation resources.
-
Have more questions?
- Join our Discord server, ask us on #sea-vl or #discussion-forum, and we’ll be happy to help!
