BREAKING
NVIDIA Unveils Flex-Forcing
Two paradigms, one model
Bidirectional diffusion
●
Attends to all frames
●
High global coherence
●
Costly and slow
Autoregressive
●
Generates frames in sequence
●
Fast streaming, long clips
●
Prone to error drift
Flexible chunking at inference
1
Bidirectional
↓
2
Hybrid
↓
3
Autoregressive
Spotlight at ICML 2026
Gains on VBench and stability
Real-world testing still ahead
AI NEWS BLITZ
NVIDIA has released Flex-Forcing, a video model that switches generation modes at inference.