diff --git a/README.md b/README.md
index e743d6b..873758b 100644
--- a/README.md
+++ b/README.md
@@ -2,6 +2,14 @@ GNet-pose
 =======
 Project Page: [http://guanghan.info/projects/guided-fractal/](http://guanghan.info/projects/Guided-Fractal/)
 
+UPDATE 9/27/2018:
+
+Prototxts and model that achieved 93.9Pck on LSP dataset. 
+[http://guanghan.info/download/Data/GNet_update.zip](http://guanghan.info/download/Data/GNet_update.zip)
+
+When I was replying e-mails, it occurred to me that the models that I had uploaded was around May/June 2017 (performance in old arxiv version), and in August 2017 the performance was improved to 93.9 on LSP with a newer caffe version which fixed the downsampling and/or upsampling deprecation problem (Yeah, it "magically" improved the performance).
+The best model was 94.0071 on LSP dataset, but it was not uploaded nor published on the benchmark. 
+
 --------
 ## Overview
 Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation.
@@ -104,15 +112,16 @@ GNet-pose is released under the Apache License Version 2.0 (refer to the LICENSE
 
 ---
 ## Citation
-The details are published as a technical report on arXiv. If you use the code and models, please cite the following paper:
-[arXiv:1705.02407](http://arxiv.org/abs/1705.02407).
-
-	@article{ning2017knowledge,
-	  title={Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation},
-	  author={Ning, Guanghan and Zhang, Zhi and He, Zhihai},
-	  journal={arXiv preprint arXiv:1705.02407},
-	  year={2017}
-	}
+If you use the code and models, please cite the following paper:
+[TMM 2017](http://ieeexplore.ieee.org/document/8064661/).
+
+	@article{ning2017knowledge, 
+	 author={G. Ning and Z. Zhang and Z. He}, 
+         journal={IEEE Transactions on Multimedia}, 
+         title={Knowledge-Guided Deep Fractal Neural Networks for Human Pose Estimation}, 
+         year={2017}, 
+         doi={10.1109/TMM.2017.2762010}, 
+         ISSN={1520-9210}, }
 
 
 ---