Research map / Video / Video generation
SALSA-V: Shortcut-Augmented Long-form Synchronized Audio from Videos
Amir Dellali, Luca A Lanzendörfer, Florian Grötschla, Roger Wattenhofer
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 |
|---|---|---|---|
| train with shortcut objective instead of flow matching (ablation) | SALSA-V 643M · 32 sampling steps | DeSync: 0.499 -> 0.497 | No clear effect |
| train with shortcut objective instead of flow matching (ablation) | SALSA-V 347M · 32 sampling steps | DeSync: 0.504 -> 0.501 | No clear effect |
| train with shortcut objective instead of flow matching (ablation) | SALSA-V 643M · 32 sampling steps | FAD_VGG: 1.10 -> 1.07 | Improved |
| scale model from 347M to 643M parameters (ablation) | SALSA-V (shortcut) · 32 sampling steps | ImageBind score: 31.09 -> 33.76 | Improved |
| scale model from 643M to 1.01B parameters (ablation) | SALSA-V (shortcut) · 32 sampling steps | DeSync: 0.497 -> 0.486 | Improved |
| scale model from 643M to 1.01B parameters (ablation) | SALSA-V (shortcut) · 32 sampling steps | ImageBind score: 33.76 -> 33.69 | No clear effect |
| SALSA-V instead of Frieren reflow at 4 sampling steps | few-step sampling · 4 sampling steps | DeSync: 0.917 -> 0.536 | Improved |
| SALSA-V instead of Frieren reflow at 4 sampling steps | few-step sampling · 4 sampling steps | FAD_VGG: 2.17 -> 1.19 | Improved |
| SALSA-V instead of LoVA for 30 s generation (outpainting) | long-form generation · 30-second long-form samples | DeSync: 1.225 -> 0.526 | Improved |
| SALSA-V instead of MMAudio | human study · Human study MOS | Synchronization: 2.92 -> 3.52 | Improved |
| SALSA-V instead of MMAudio | human study · Human study MOS | Quality: 3.16 -> 2.96 | Hurt |
| reduce sampling from 32 to 4 steps with shortcut model | SALSA-V · Generation time, 10 s clip | time (s): 4.84 -> 0.71 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- AudioX — AudioX: A Unified Framework for Anything-to-Audio Generation (ICLR 2026) · better on 2 of 3 shared results
- FoleyCrafter — FoleyCrafter: Bring Silent Videos to Life with Lifelike and Synchronized Sounds (IJCV 2026) · better on 2 of 2 shared results
- Frieren — Frieren: Efficient Video-to-Audio Generation Network with Rectified Flow Matching (NeurIPS 2024) · better on 2 of 2 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
- MMAudio: Taming Multimodal Joint Training for High-Quality Video-to-Audio Synthesis (CVPR 2025 · similarity 0.89)
- Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows (CVPR 2025 · similarity 0.88)
- Audio-Sync Video Generation with Multi-Stream Temporal Control (NeurIPS 2025 · similarity 0.88)
- Synchronized Video-to-Audio Generation via Mel Quantization-Continuum Decomposition (CVPR 2025 · similarity 0.88)
- Diverse and Aligned Audio-to-Video Generation via Text-to-Video Model Adaptation (AAAI 2024 · similarity 0.88)
- Model-Guided Dual-Role Alignment for High-Fidelity Open-Domain Video-to-Audio Generation (NeurIPS 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.
