Skip to content

Token pruning

Token pruning removes low-salience video tokens in selected steps. It targets model-level redundancy in long spatiotemporal sequences.

The paper treats token pruning as a sequence-reduction family: the token-pruning agent tunes the pruning criterion, pruning ratio, layer schedule, timestep schedule, and reconstruction rule, then validates temporal coherence and late-step fidelity.

In Sol-Engine

LTX-2.3 uses token pruning in the full optimization stack together with kernel fusion, cache, PISA sparse attention, and NVFP4.

Implemented entries:

  • python/sglang/multimodal_gen/runtime/efficiency/techniques/token_prune.py
  • python/sglang/multimodal_gen/runtime/pipelines_core/stages/ltx_2_denoising.py
  • python/sglang/multimodal_gen/runtime/efficiency/models/ltx2_spec.py
  • python/sglang/multimodal_gen/runtime/efficiency/spec.py

Methods

Method Open-source status Integration Role in the design space
Feature-norm pruning open-source local local LTX runtime Sol-Engine LTX-2.3 midpoint token pruning implementation
ToMe-SD open-source external adapter diffusion token merging baseline family

Practical notes

  • Prune only in steps where visual sensitivity is acceptable.
  • Keep prompt, seed, scheduler, and resolution fixed when comparing variants.
  • Inspect videos side by side; scalar metrics are not enough for final acceptance.