Feature-norm pruning¶
Feature-norm pruning is the Sol-Engine token-pruning implementation used for LTX-2.3. It scores video tokens by feature norm, keeps the highest-salience tokens, runs the expensive block stack on that subset, and compensates dropped tokens from prior state.
Sol-Engine placement¶
Sol-Engine uses feature-norm pruning in LTX-2.3 stage-2 midpoint steps together with PISA, kernel fusion, cache, and NVFP4.
Tunable knobs¶
- keep ratio.
- pruning steps.
- salience metric.
- compensation rule.
- prunable token seam.
Validation¶
Middle-step pruning should preserve temporal coherence and late-step detail. Do not accept it from latency alone.