Research map / Video / Video generation
CAFA: a Controllable Automatic Foley Artist
Roi Benita, Michael Finkelson, Tavi Halperin, Gleb Sterkin, Yossi Adi
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 CLIP visual features alongside AV-CLIP (ablation) | CAFA · VGGSound test | FAD / DeSync: FAD 12.60 -> 14.44; DeSync 0.96 -> 1.02 | No clear effect |
| add fine-tuning on VisualSound after VGGSound training (ablation) | CAFA-B · VGGSound test | IS / DeSync: IS 11.84 -> 13.45; DeSync 1.00 -> 0.96 | Improved |
| swap Stable-Audio-Open for TangoFlux as T2A backbone (ablation) | CAFA · VGGSound test | FAD: FAD 12.60 -> 19.94; IS 13.45 -> 16.94 | Mixed |
| use asymmetric CFG scale between 0.5 and 1 (ablation) | CAFA · VGGSound-Sparse | overall trade-off | Improved |
| add text conditioning to MMAudio | MMAudio · VGGSound test | FAD | Hurt |
| use CAFA instead of MMAudio when text conflicts with video | text-video-to-audio · VGGSound-Sparse conflicting captions | Acc: 0.36 -> 0.87 | Improved |
| use CAFA instead of MMAudio when text aligns with video | text-video-to-audio · VGGSound test | FAD: 6.89 -> 12.60 | Hurt |
| adapter-based training on frozen T2A model instead of full training | CAFA · training cost | GPU hours: 304 -> ~48 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- VATT — VATT: Transformers for Multimodal Self-Supervised Learning from Raw Video, Audio and Text (NeurIPS 2021) · better on 0 of 1 shared result
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
- Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows (CVPR 2025 · similarity 0.88)
- Diverse and Aligned Audio-to-Video Generation via Text-to-Video Model Adaptation (AAAI 2024 · similarity 0.86)
- Read, Watch and Scream! Sound Generation from Text and Video (AAAI 2025 · similarity 0.86)
- Tri-Ergon: Fine-Grained Video-to-Audio Generation with Multi-Modal Conditions and LUFS Control (AAAI 2025 · similarity 0.85)
- ThinkSound: Chain-of-Thought Reasoning in Multimodal LLMs for Audio Generation and Editing (NeurIPS 2025 · similarity 0.85)
- Audio-Sync Video Generation with Multi-Stream Temporal Control (NeurIPS 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.
