Research map / Video / Video segmentation and tracking
Characterizing and Improving Stability in Neural Style Transfer
Agrim Gupta, Justin Johnson, Alexandre Alahi, Li Fei-Fei
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 |
|---|---|---|---|
| replace per-frame feedforward stylization with a recurrent network trained with optical-flow temporal consistency loss | Real-Time baseline · DAVIS stability, Composition XIV | SSIM between stylized adjacent patches: 0.31 -> 0.42 | Improved |
| add recurrent input and temporal consistency loss | Real-Time baseline · DAVIS stability, Metzinger | SSIM between stylized adjacent patches: 0.31 -> 0.42 | Improved |
| add recurrent input and temporal consistency loss | Real-Time baseline · DAVIS stability, Rain Princess | SSIM between stylized adjacent patches: 0.41 -> 0.49 | Improved |
| add recurrent input and temporal consistency loss | Real-Time baseline · DAVIS stability, The Wave | PSNR between stylized adjacent patches: 24.3 -> 24.8 | Improved |
| add recurrent input and temporal consistency loss | Real-Time baseline · DAVIS stability, Mosaic | SSIM between stylized adjacent patches: 0.31 -> 0.39 | Improved |
| use feedforward recurrent network instead of per-frame optimization | Optim baseline · DAVIS stability, The Wave | PSNR between stylized adjacent patches: 25.5 -> 24.8 | Hurt |
| use feedforward recurrent network instead of per-frame optimization | Optim baseline · DAVIS stability, SSIM across five styles | SSIM between stylized adjacent patches: 0.48 -> 0.54 | Improved |
| use feedforward network instead of optimization | Optim baseline · Per-frame runtime, 1024 x 1024 | seconds per frame: 199.6 -> 0.141 | Improved |
| use feedforward network instead of optimization | Optim baseline · Per-frame runtime, 256 x 256 | seconds per frame: 22.14 -> 0.024 | Improved |
| add recurrent temporal-consistency training versus per-frame network | Real-Time baseline · Per-frame runtime, 512 x 512 | seconds per frame: 0.044 -> 0.044 | No clear effect |
| compare ours to Real-Time baseline in user study | Real-Time baseline · User study (AMT), More Flicker | videos judged to flicker more: 26 vs 193 | Improved |
| compare ours to Real-Time baseline in user study | Real-Time baseline · User study (AMT), Overall Prefer | videos preferred: 133 vs 80 | Improved |
Closest papers by meaning
- Real-Time Neural Style Transfer for Videos (CVPR 2017 · similarity 0.94)
- Adjustable Real-time Style Transfer (ICLR 2020 · similarity 0.91)
- Consistent Video Style Transfer via Compound Regularization (AAAI 2020 · similarity 0.91)
- A Flexible Convolutional Solver for Fast Style Transfers (CVPR 2019 · similarity 0.90)
- Coherent Online Video Style Transfer (ICCV 2017 · similarity 0.90)
- Learning Linear Transformations for Fast Image and Video Style Transfer (CVPR 2019 · similarity 0.90)
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.
