X-Decoder
A generalized decoder that predicts segmentation masks and language tokens in one shared semantic space.
About
X-Decoder from Microsoft Research treats segmentation and vision-language tasks as the same decoding problem. It takes generic non-semantic queries plus semantic queries derived from text, and decodes both pixel-level masks and language tokens in one space. After pretraining on a modest amount of segmentation data plus millions of image-text pairs, the same model does open-vocabulary segmentation, referring segmentation, captioning, and VQA without task-specific heads.
Fine-tuned X-Decoder-L set a new state of the art on ADE20K panoptic segmentation (51.0 PQ) at publication, and the base Focal-L model reports 56.5 COCO PQ, 46.7 mAP, and 67.2 mIoU. It was the conceptual predecessor to SEEM, and its checkpoints live on Hugging Face under the xdecoder org. License status was not explicitly confirmed in the sources checked, so treat commercial use as unverified.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported: 51.0 PQ ADE20K panoptic (finetuned L, SOTA at publication); Focal-L 56.5 COCO PQ / 46.7 mAP / 67.2 mIoU
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (metrics tier, metrics-v1). Evidence: CVPR 2023 paper (arXiv:2212.11270) with ADE20K and COCO numbers confirmed via the paper and the official repo checkpoint table. Strengths: unified pixel and language decoding, strong transferability, predecessor to SEEM. Docked for: (1) repo license not explicitly confirmed in the sources checked; (2) superseded in mindshare by SEEM and later universal segmenters. License: not explicitly confirmed — flag for review.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · CVPR 2023 (arXiv 2212.11270, Dec 2022)
- · 51.0 PQ ADE20K panoptic (finetuned L, SOTA at publication), author-reported
- · Conceptual predecessor to SEEM; checkpoints on Hugging Face (xdecoder/X-Decoder)