Research map / Video / Video generation
SummDiff: Generative Modeling of Video Summarization with Diffusion
Kwanseok Kim, Jaehoon Hahm, Sumin Kim, Jinhwan Sul, Byunghak Kim, Joonseok Lee
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 |
|---|---|---|---|
| add video importance denoiser with summed query to encoder (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.071 -> 0.079 | Improved |
| replace fixed random score embeddings with learnable ones (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.079 -> 0.086 | Improved |
| add AdaLN on time embedding (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.086 -> 0.145 | Improved |
| add AdaLN on time and positional encodings (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.145 -> 0.171 | Improved |
| add score quantization into codebook embeddings (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.125 -> 0.171 | Improved |
| add classifier-free guidance (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.171 -> 0.175 | Improved |
| add self-attention guidance (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.175 -> 0.177 | Improved |
| increase DDIM steps from 1 to 100 (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.170 -> 0.182 | Improved |
| increase DDIM steps from 1 to 10 (inference time) (ablation) | SummDiff · Mr.HiSum | inference time (ms): 11.02 -> 49.73 | Hurt |
| increase quantization bins K from 5 to 200 (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.145 -> 0.175 | Improved |
| increase quantization bins K from 200 to 800 (ablation) | SummDiff · Mr.HiSum | Kendall tau: 0.175 -> 0.172 | No clear effect |
| train on individual annotator scores instead of aggregated (all annotations) (ablation) | SummDiff · SumMe | Kendall tau: 0.176 -> 0.256 | Improved |
Later papers that compared with this method
1 later accepted paper reports SummDiff as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- TRINITY: A Multi-Perspective Benchmark for Personal-Style Video Highlight Detection (ECCV 2026 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- CSTA — CSTA: CNN-based Spatiotemporal Attention for Video Summarization (CVPR 2024) · better on 14 of 15 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Video Summarization by Learning From Unpaired Data (CVPR 2019 · similarity 0.91)
- Video Summarization Using Denoising Diffusion Probabilistic Model (AAAI 2025 · similarity 0.91)
- Query-Focused Video Summarization: Dataset, Evaluation, and a Memory Network Based Approach (CVPR 2017 · similarity 0.90)
- Video Summarization with Long Short-Term Memory (ECCV 2016 · similarity 0.90)
- Deep Reinforcement Learning for Unsupervised Video Summarization With Diversity-Representativeness Reward (AAAI 2018 · similarity 0.89)
- Video Summarization with Large Language Models (CVPR 2025 · similarity 0.88)
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.
