---
pipeline_tag: image-segmentation
license: cc-by-nc-4.0
base_model: facebook/maskformer-swin-tiny-ade
library_name: zeromodels
tags:
- keras
- zeromodels
- maskformer
- universal-segmentation
- image-segmentation
- arxiv:2107.06278
- pytorch
- jax
- tf
---

## ***See [our collection](https://huggingface.co/collections/zeromodels/maskformer-6a8eaf6a43e1a5079d6cc8ef) for all versions of MaskFormer.***

# Run MaskFormer with Keras 3: JAX, PyTorch, or TensorFlow

[![GitHub](https://img.shields.io/badge/GitHub-ZeroModels-black?logo=github)](https://github.com/IMvision12/ZeroModels) [![Docs](https://img.shields.io/badge/Docs-MaskFormer-blue)](https://imvision12.github.io/ZeroModels/maskformer/) [![Collection](https://img.shields.io/badge/HF-MaskFormer%20collection-yellow)](https://huggingface.co/collections/zeromodels/maskformer-6a8eaf6a43e1a5079d6cc8ef)

# zeromodels/maskformer-swin-tiny-ade

Paper: [Per-Pixel Classification is Not All You Need for Semantic Segmentation (arXiv:2107.06278)](https://arxiv.org/abs/2107.06278) · [HF Papers](https://huggingface.co/papers/2107.06278)

MaskFormer reframes segmentation as mask classification: a backbone and pixel decoder feed a transformer decoder whose queries each predict a binary mask and a class. One architecture covers semantic, instance, and panoptic outputs via post-processing.

For more details on the model, please go to the upstream [model card](https://huggingface.co/facebook/maskformer-swin-tiny-ade).

Pure-**Keras 3** conversion of [`facebook/maskformer-swin-tiny-ade`](https://huggingface.co/facebook/maskformer-swin-tiny-ade) for [zeromodels](https://github.com/IMvision12/ZeroModels). One implementation runs unmodified on **TensorFlow / Torch / JAX**.

This is a **universal segmentation** checkpoint (`MaskFormerUniversalSegment`) trained on ADE20K.

## ✨ Quick start

```python
import os
os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.maskformer import MaskFormerUniversalSegment, MaskFormerImageProcessor

model = MaskFormerUniversalSegment.from_weights("zeromodels/maskformer-swin-tiny-ade")
processor = MaskFormerImageProcessor.from_weights("zeromodels/maskformer-swin-tiny-ade")

image = Image.open("your_image.jpg").convert("RGB")
output = model(processor(image)["pixel_values"], training=False)
result = processor.post_process_panoptic_segmentation(
    output, target_size=(image.height, image.width)
)
print(result["segmentation"].shape)
```

Load any MaskFormer variant the same way with `from_weights("zeromodels/<variant>")`:

| Variant | Hub | Dataset |
|---|---|---|
| `maskformer-swin-tiny-coco` | [`zeromodels/maskformer-swin-tiny-coco`](https://huggingface.co/zeromodels/maskformer-swin-tiny-coco) | COCO |
| `maskformer-swin-small-coco` | [`zeromodels/maskformer-swin-small-coco`](https://huggingface.co/zeromodels/maskformer-swin-small-coco) | COCO |
| `maskformer-swin-base-coco` | [`zeromodels/maskformer-swin-base-coco`](https://huggingface.co/zeromodels/maskformer-swin-base-coco) | COCO |
| `maskformer-swin-tiny-ade` | [`zeromodels/maskformer-swin-tiny-ade`](https://huggingface.co/zeromodels/maskformer-swin-tiny-ade) | ADE20K |
| `maskformer-swin-base-ade` | [`zeromodels/maskformer-swin-base-ade`](https://huggingface.co/zeromodels/maskformer-swin-base-ade) | ADE20K |

## Tips

- Set `KERAS_BACKEND` **before** importing Keras / zeromodels.
- Prefer `MaskFormerImageProcessor.from_weights(...)` so resolution matches the variant.
- See [MaskFormer docs]({DOCS_URL}) and [Loading Weights](https://imvision12.github.io/ZeroModels/loading_weights/).
- Community / upstream weights: `MaskFormerUniversalSegment.from_weights("hf:facebook/maskformer-swin-tiny-ade")`.

## Special Thanks

A huge thank you to the Facebook AI Research MaskFormer authors for creating and releasing these models.

License: CC-BY-NC-4.0 (non-commercial).