See our collection for all versions of EoMT.

Run EoMT with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/eomt_large_coco_panoptic_640

Paper: Your ViT is Secretly an Image Segmentation Model (arXiv:2503.19108) · HF Papers

EoMT (Encoder-only Mask Transformer) keeps segmentation inside a plain ViT: learned query tokens are concatenated with patch tokens and run through the same ViT blocks. No pixel decoder, no deformable attention decoder.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of tue-mps/coco_panoptic_eomt_large_640 for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is a panoptic checkpoint (EoMTUniversalSegment).

✨ Quick start

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

from PIL import Image
from zeromodels.models.eomt import EoMTUniversalSegment, EoMTImageProcessor

model = EoMTUniversalSegment.from_weights("zeromodels/eomt_large_coco_panoptic_640")
processor = EoMTImageProcessor.from_weights("zeromodels/eomt_large_coco_panoptic_640")

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 EoMT variant the same way with from_weights("zeromodels/<variant>"):

Tips

• Set KERAS_BACKEND before importing Keras / zeromodels.
• Use the post-processor that matches the checkpoint task (panoptic / instance / semantic).
• See EoMT docs and Loading Weights.
• Community / upstream weights: EoMTUniversalSegment.from_weights("hf:tue-mps/coco_panoptic_eomt_large_640").

Special Thanks

A huge thank you to the TU/e MPS EoMT authors for creating and releasing these models.

License: MIT.