Research map / Image generation / Autoregressive and latent generation
Low Distortion Block-Resampling with Spatially Stochastic Networks
Sarah Hong, Martin Arjovsky, Darryl Barnhart, Ian Thompson
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 spatial latent blocks and block resampling instead of a global latent | SSN versus StyleGAN2 · FFHQ 256x256 | FID: 19.76 -> 12.24 | Improved |
| use spatial latent blocks and block resampling instead of a global latent | SSN versus StyleGAN2 · LSUN Churches 256x256 | FID: 3.65 -> 8.68 | Hurt |
| add low-distortion regularization with λ_D=100 rather than no distortion regularization (ablation) | SSN · FFHQ 256x256 | Distortion: 0.028 -> 0.0043 | Improved |
| add low-distortion regularization with λ_D=100 rather than no distortion regularization (ablation) | SSN · FFHQ 256x256 | FID: 12.24 -> 15.24 | Hurt |
| set λ_D to 1 instead of 0 (ablation) | SSN · FFHQ 256x256 | Distortion: 0.028 -> 0.028 | No clear effect |
| set λ_D to 10 instead of 0 (ablation) | SSN · FFHQ 256x256 | Distortion: 0.028 -> 0.017 | Improved |
| set λ_D to 1000 instead of 0 (ablation) | SSN · FFHQ 256x256 | FID: 12.24 -> 66.74 | Hurt |
| set λ_D to 10000 instead of 0 (ablation) | SSN · FFHQ 256x256 | FID: 12.24 -> 128.99 | Hurt |
| replace StyleGAN2 path length regularization with the proposed low-distortion path length term (ablation) | SSN · FFHQ 256x256 | Distortion | No clear effect |
| replace SPADE normalization with spatially modulated convolution (ablation) | spatially conditioned generator · Qualitative spatial conditioning | droplet artifacts | Improved |
Closest papers by meaning
- Generative Refinement Network for Visual Synthesis (ECCV 2026 · similarity 0.87)
- Consistency Posterior Sampling for Diverse Image Synthesis (CVPR 2025 · similarity 0.86)
- GENERATING HIGH FIDELITY IMAGES WITH SUBSCALE PIXEL NETWORKS AND MULTIDIMENSIONAL UPSCALING (ICLR 2019 · similarity 0.86)
- Plug & Play Generative Networks: Conditional Iterative Generation of Images in Latent Space (CVPR 2017 · similarity 0.86)
- Restoration by Generation with Constrained Priors (CVPR 2024 · similarity 0.85)
- Stochastic Forward–Backward Deconvolution: Training Diffusion Models with Finite Noisy Datasets (ICML 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.
