metadata
title: README
emoji: 🐠
colorFrom: pink
colorTo: green
sdk: static
pinned: false
fjuice
Yoooo 👋 welcome to fjuice.
We build vision systems, datasets, and models that don’t take themselves too seriously — but still run like production-grade tools.
Think:
- synthetic data pipelines that actually scale
- open-vocab vision models that don’t fall apart in the wild
- APIs that feel like:
fjuice.detect()instead of corporate SDK soup
🍊 What is fjuice?
fjuice is a research + engineering project focused on:
- synthetic dataset generation at scale (Juicebox series)
- open-vocab vision models (JuiceJet series)
- segmentation + grounding systems
- messy-real-world robustness via controlled randomness
We like structured chaos.
🧃 Ecosystem
🧱 Juicebox
Synthetic datasets for vision training.
- multi-million to multi-billion scale generation
- heavy augmentation + realism injection
- segmentation-first design
- open vocab friendly
✈️ JuiceJet
Vision models trained on Juicebox datasets.
- segmentation / detection / grounding models
- open-vocab reasoning over visual scenes
- designed for robustness, not just benchmarks
⚙️ fjuice core
A lightweight API layer for running models:
fjuice.detect(image)
fjuice.segment(image)
fjuice.infer(prompt, image)
Yes, it’s intentionally simple.
🧠 Philosophy
We don’t optimize for:
- fancy corporate frameworks
- over-engineered abstractions
- benchmark-only performance
We optimize for:
- real-world robustness
- dataset diversity
- reproducible chaos
- fast iteration
🚀 Why it exists
Because most vision datasets are either:
- too clean
- too small
- too static
- or too boring
So we built something messy enough to generalize.
🧪 Status
Actively evolving.
Stuff will break. Stuff will change. That’s the point.
🪪 License
See LICENSE file.
🌐 Links
- Hugging Face Org: (Visit Here!)
- Models: JuiceJet series
- Datasets: Juicebox series
- CONTRIBUTING: https://tally.so/r/VLN9zN
Made with chaos, caffeine, and too many GPUs.