SegFormer
NVIDIASegFormer is a semantic segmentation model line from NVIDIA Research that pairs a hierarchical Transformer encoder, in sizes B0 to B5, with a lightweight all-MLP decoder. The encoders are pretrained on ImageNet-1K and the released checkpoints are fine-tuned on ADE20K and Cityscapes, with training and evaluation code built on MMSegmentation. NVIDIA's checkpoints are also on the Hugging Face Hub in Transformers format.
Openness
2 high confidence- weights
- open (B0 to B5 ADE20K and Cityscapes checkpoints on the Hub, ungated)
- data
- open (ImageNet-1K for the encoders, then ADE20K or Cityscapes, all public benchmarks)
- code
- open (training and evaluation code and configs, built on MMSegmentation)
- license
- NVIDIA-Source-Code-License (code and checkpoints; research or evaluation use only, commercial use reserved to NVIDIA)
SegFormer's code and checkpoints are under the NVIDIA Source Code License, which allows use only for research or evaluation and reserves commercial use to NVIDIA. The public training data and the published training code do not lift that restriction.
- https://huggingface.co/api/models?author=nvidia&search=segformer&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
Fifteen nvidia SegFormer repositories, all gated false and tagged license other.
- https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512/raw/main/README.md recorded 2026-11-12
segformer-b0-finetuned-ade-512-512 card: license other, "The license for this model can be found here" linking the repository LICENSE; the encoder "is first pre-trained on ImageNet-1k" and then fine-tuned on a downstream dataset.
- https://raw.githubusercontent.com/NVlabs/SegFormer/master/LICENSE recorded 2026-11-12
NVIDIA Source Code License for SegFormer, 3.3 Use Limitation: "The Work and any derivative works thereof only may be used or intended for use non-commercially"; "non-commercially" means "for research or evaluation purposes only".
- https://raw.githubusercontent.com/NVlabs/SegFormer/master/README.md recorded 2026-11-12
README: "This repository contains the official Pytorch implementation of training & evaluation code and the pretrained models"; training with tools/train.py on ADE20K configs from encoders "pretrained on ImageNet-1K".
Adoption
3 high confidenceHugging Face downloads over the trailing 30 days for the five declared checkpoints, most of them for the smallest ADE20K model. The other NVIDIA SegFormer checkpoints add comparatively little.
- https://huggingface.co/api/models?author=nvidia&search=segformer&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
segformer-b0-finetuned-ade-512-512 276,662, b2-ade 157,213, b3-ade 136,827, b5-ade-640 57,233 and b5-cityscapes-1024 51,561 downloads in the trailing 30 days (679,496 together); the fifteen repositories sum to 802,986.
Capability
3 high confidenceEach SegFormer checkpoint labels pixels with the classes of the benchmark it was fine-tuned on, so a new set of classes needs a training run, as with RT-DETR's COCO detectors.
- https://raw.githubusercontent.com/NVlabs/SegFormer/master/README.md recorded 2026-11-12
README: SegFormer-B0 to B5 results and trained weights for ADE20K and Cityscapes; single- and multi-GPU training commands with the local_configs/segformer configs.
Verified 2026-11-12