MetaStyle: Three-Way Trade-off among Speed, Flexibility, and Quality in Neural Style Transfer

Authors: Chi Zhang, Yixin Zhu, Song-Chun Zhu1254-1261

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The qualitative and quantitative analysis in the experiments demonstrates that the proposed approach achieves high-quality arbitrary artistic style transfer effectively, with a good trade-off among speed, flexibility, and quality. Comprehensive experiments with both qualitative and quantitative analysis, compared with prior neural style transfer methods, demonstrate that the proposed method achieves a good trade-off among speed, flexibility, and quality. 5 Experiments
Researcher Affiliation Academia Chi Zhang, Yixin Zhu, Song-Chun Zhu International Center for AI and Robot Autonomy (CARA) {chizhang,yzhu,sczhu}@cara.ai
Pseudocode Yes Algorithm 1: Meta Style
Open Source Code No The paper mentions "additional videos are provided in the supplementary files," but does not provide an explicit statement or a link indicating that the source code for the described methodology is publicly available.
Open Datasets Yes We train our model using MS-COCO (?) as our content dataset and Wiki Art test set (?) as our style dataset. (Citations: Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll ar, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In Proceedings of European Conference on Computer Vision (ECCV). Nichol, K. 2016. Painter by numbers, wikiart.)
Dataset Splits Yes The inner objective uses a model initialized with θ and only optimizes contents in the training set, whereas the outer objective tries to generalize to contents in the validation set. The explicit training-validation separation in the framework forces the style transfer model to generalize to unobserved content images without over-fitting to the training set. Input : content training dataset Dtr, content validation dataset Dval, style dataset Dstyle, inner learning rate δ, outer learning rate η, number of inner updates T
Hardware Specification Yes The entire model is trained on a Nvidia Titan Xp with only 0.1 million iterations. which amounts to an additional 24 seconds of training time with a Titan Xp GPU.
Software Dependencies No The paper mentions optimizers like Adam and SGD but does not specify any software libraries or their version numbers (e.g., PyTorch 1.x, TensorFlow 2.x) that were used for implementation.
Experiment Setup Yes We use Adam (?) with a learning rate 0.001 to optimize the outer objective and vanilla SGD with a learning rate 0.0001 for the inner objective. All batches are of size 4. We fix α = 1, β = 1 105 across all the experiments. Content loss is computed on relu2_2 of a pre-trained VGG16 model and style loss over relu1_2, relu2_2, relu3_3 and relu4_3. To encourage fast adaptation, we constrain T = 1. The entire model is trained on a Nvidia Titan Xp with only 0.1 million iterations.