LTX-2.3¶
LTX-2.3 is the 22B high-resolution pipeline. The optimized path combines kernel fusion, cache, PISA sparse attention, NVFP4, and token pruning.
Speed¶
Full optimization uses kernel fusion + cache + PISA + NVFP4 + token-prune and reaches 2.40x speedup. Measured on GB200 with 1088x1920, 241 frames, warmup excluded.
Launch¶
bash scripts/ltx/run_ltx23_sglang_hq_1080p10s.sh baseline
bash scripts/ltx/run_ltx23_sglang_hq_1080p10s.sh fullopt
fullopt is self-contained. Override MODEL_PATH, DISTILLED_LORA, or SPATIAL_UPSAMPLER when weights live outside the default cache location.
Techniques¶
- Kernel fusion: kernel fusion reduces DiT operator overhead.
- Cache: stage-specific reuse avoids redundant work.
- Sparse attention: PISA targets redundant video self-attention.
- Quantization: NVFP4 targets video FFN compute.
- Token pruning: low-salience tokens are removed in selected refine steps.