Research map / Image generation / Autoregressive and latent generation
A Powerful Generative Model Using Random Weights for the Deep Image Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace trained VGG weights with random Gaussian weights (sigma 0.015) for representation inversion (ablation) | VGG-19 inversion · ILSVRC 2012 validation images | reconstruction quality and convergence speed: 2000 -> 100 iterations | Helped |
| Double the iteration limit to 4000 on the pretrained VGG (ablation) | pretrained VGG inversion · monkey and house images | reconstruction: 2000 -> 4000 | No clear effect |
| Use stacked random weight network instead of purely random weights (ablation) | ranVGG · girl image inversion | reconstruction distance and variation | Helped |
| Use 19-layer random VGG with average pooling instead of AlexNet with max pooling (Mahendran et al.) | inversion at pool3 and pool4 · image inversion | perceptual quality | Helped |
| Add conv4_1 to texture matching layers (conv1_1 to conv3_1) (ablation) | ranVGG texture synthesis · natural textures | texture quality | Helped |
| Compare random ranVGG against trained VGG for texture synthesis | texture synthesis · regular textures such as cargo | texture quality | Hurt |
| Use untrained ranVGG instead of pretrained VGG for style transfer | style transfer · five famous artworks | artistic image quality | No clear effect |
| Use pretrained VGG for inversion from pool5 | pretrained VGG · image inversion | contour recovery | Hurt |
Closest papers by meaning
- Images as Weight Matrices: Sequential Image Generation Through Synaptic Learning Rules (ICLR 2023 · similarity 0.88)
- Robust and Generalizable Visual Representation Learning via Random Convolutions (ICLR 2021 · similarity 0.88)
- Texture Networks: Feed-forward Synthesis of Textures and Stylized Images (ICML 2016 · similarity 0.88)
- Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space (CVPR 2017 · similarity 0.87)
- A Learned Representation For Artistic Style (ICLR 2017 · similarity 0.87)
- Restyling Unsupervised Concept Based Interpretable Networks with Generative Models (ICLR 2025 · similarity 0.85)
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.