Research map / Video / Video generation
OmniForcing: Unleashing Real-time Joint Audio-Visual Generation
Yaofeng Su, Yuming Li, Zeyue Xue, Jie Huang, Siming Fu, Haoran Li, Haoyang Huang, Nan Duan
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 Identity RoPE on sink tokens instead of incremental positions (ablation) | OmniForcing Stage II · Stage II stabilization ablation | loss (sigma=0.5): 0.402 vs. 0.081 | Improved |
| add audio sink tokens (S>=4) versus no stabilizer (ablation) | OmniForcing Stage II · Stage II stabilization ablation | convergence: No Stabilizer NaN; S=4 loss 0.141 | Improved |
| use S=1 or S=2 sink tokens instead of S>=4 (ablation) | OmniForcing Stage II · Stage II stabilization ablation | convergence: NaN gradients at S=2 and S=1 | Hurt |
| increase sink pool from 4 to 16 tokens (ablation) | OmniForcing Stage II · Stage II stabilization ablation | loss (sigma=0.5): 0.141 -> 0.081 | Improved |
| increase sink pool from 16 to 24 tokens (ablation) | OmniForcing Stage II · Stage II stabilization ablation | loss (sigma=0.5): 0.081 -> 0.110 | Hurt |
| replace QK-Norm with audio sink tokens (ablation) | OmniForcing Stage II · Stage II stabilization ablation | loss (sigma=0.5): 0.232 -> 0.081 | Improved |
| replace Tanh-Gated Attention with audio sink tokens (ablation) | OmniForcing Stage II · Stage II stabilization ablation | loss (sigma=0.5): 1.258 -> 0.081 | Improved |
| increase sink pool from 8 to 16 tokens (ablation) | OmniForcing Stage II · Stage II stabilization ablation | max gradient norm: 21.95 -> 9.23 | Improved |
| distill bidirectional LTX-2 into streaming OmniForcing | LTX-2 19B · JavisBench | runtime: 197s -> 5.7s | Improved |
| distill bidirectional LTX-2 into streaming OmniForcing | LTX-2 19B · JavisBench | JavisScore: 0.253 -> 0.208 | Hurt |
| distill bidirectional LTX-2 into streaming OmniForcing | LTX-2 19B · VBench (same prompts) | aesthetic quality: 0.569 -> 0.595 | Improved |
| distill bidirectional LTX-2 into streaming OmniForcing | LTX-2 19B · JavisBench | FVD: 125.4 -> 137.2 | Hurt |
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.
- JavisDiT++ — JavisDiT: Joint Audio-Video Diffusion Transformer with Hierarchical Spatio-Temporal Prior Synchronization (ICLR 2026) · better on 3 of 4 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
- Seeing and Hearing: Open-domain Visual-Audio Generation with Diffusion Latent Aligners (CVPR 2024 · similarity 0.89)
- UniFlow-Audio: Unified Flow Matching for Audio Generation from Omni-Modalities (NeurIPS 2026 · similarity 0.87)
- Native Audio-Visual Alignment for Generation (NeurIPS 2026 · similarity 0.87)
- Foley-Flow: Coordinated Video-to-Audio Generation with Masked Audio-Visual Alignment and Dynamic Conditional Flows (CVPR 2025 · similarity 0.87)
- OmniTalker: One-shot Real-time Text-Driven Talking Audio-Video Generation With Multimodal Style Mimicking (NeurIPS 2025 · similarity 0.87)
- MMDisCo: Multi-Modal Discriminator-Guided Cooperative Diffusion for Joint Audio and Video Generation (ICLR 2025 · similarity 0.86)
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.
