Research map / Image generation / GANs
Precomputed Real-Time Texture Synthesis with Markovian Generative Adversarial Networks
Chuan Li, Michael Wand
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use adversarial (Markovian GAN) loss instead of pixel-level Euclidean loss for training the decoder (ablation) | MGANs decoder · Picasso style transfer, training and test photos | visual quality (qualitative) | Improved |
| compare at neural encoding level instead of pixel level (ablation) | VAE decoder · Picasso style transfer | visual quality (qualitative) | Improved |
| use pre-trained VGG_19 features for the discriminator instead of training it from pixels (ablation) | MDANs discriminator · unguided texture synthesis | visual quality (qualitative) | Improved |
| feed discriminator from lower layer relu2_1 (patch 16) instead of relu3_1 (ablation) | MDANs discriminator · unguided texture synthesis | sharpness vs structure (qualitative) | Mixed |
| feed discriminator from higher layer relu4_1 (patch 4) instead of relu3_1 (ablation) | MDANs discriminator · unguided texture synthesis | structure vs flexibility (qualitative) | Mixed |
| use larger patches (16) instead of 8 in discriminator (ablation) | MDANs discriminator · unguided texture synthesis | rigidity (qualitative) | Mixed |
| add convolutional layer to the discriminator (ablation) | MDANs discriminator · texture synthesis | synthesis quality (qualitative) | Improved |
| add further layers to the discriminator beyond one convolutional layer (ablation) | MDANs discriminator · texture synthesis | synthesis quality (qualitative) | No clear effect |
| increase discriminator channels from 4 to 64 (ablation) | MDANs discriminator · texture synthesis | synthesis quality (qualitative) | Improved |
| increase discriminator channels from 64 to 128 (ablation) | MDANs discriminator · texture synthesis | synthesis quality (qualitative) | No clear effect |
| make the generator more complex (more layers or channels) (ablation) | MGANs generator · texture synthesis | synthesis quality (qualitative) | No clear effect |
| initialize discriminator with a network pre-trained on 50 face images (ablation) | MDANs · style transfer to face images | iterations needed (qualitative): 500 iterations without -> 50 iterations with | Improved |
Closest papers by meaning
- Combining Markov Random Fields and Convolutional Neural Networks for Image Synthesis (CVPR 2016 · similarity 0.90)
- Texture Networks: Feed-forward Synthesis of Textures and Stylized Images (ICML 2016 · similarity 0.90)
- Diversified Texture Synthesis With Feed-Forward Networks (CVPR 2017 · similarity 0.90)
- GramGAN: Deep 3D Texture Synthesis From 2D Exemplars (NeurIPS 2020 · similarity 0.90)
- TextureGAN: Controlling Deep Image Synthesis With Texture Patches (CVPR 2018 · similarity 0.89)
- Fast Texture Synthesis via Pseudo Optimizer (CVPR 2020 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
