Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• OmniForcing is the first framework to distill an offline dual-stream bidirectional diffusion model into a streaming autoregressive generator for joint audio-visual generation.method 95%
• Naive causal distillation applied to dual-stream architectures triggers severe training instability due to extreme temporal asymmetry between modalities and resulting token sparsity.technical 90%
• OmniForcing addresses the information density gap through Asymmetric Block-Causal Alignment with a zero-truncation Global Prefix that prevents multi-modal synchronization drift.method 90%
• The gradient explosion caused by extreme audio token sparsity during causal shift is resolved through an Audio Sink Token mechanism with Identity RoPE constraint.method 90%
• OmniForcing uses a Joint Self-Forcing Distillation paradigm to enable dynamic self-correction of cumulative cross-modal errors from exposure bias during long rollouts.method 90%
• OmniForcing achieves state-of-the-art streaming generation at approximately 25 FPS on a single GPU.result 95%
• OmniForcing maintains multi-modal synchronization and visual quality on par with the bidirectional teacher model.result 85%
• Recent joint audio-visual diffusion models suffer from high latency due to bidirectional attention dependencies, hindering real-time applications.limitation 85%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card