Sage Attention cuts LTX 2.3 video render time 30% on RTX 3060
A ComfyUI workflow update adds Sage Attention support to LTX 2.3 video generation, delivering 30% faster performance on mid-range consumer hardware alongside transition controls and audio fixes.
A new ComfyUI workflow for LTX 2.3 video generation delivers 30% faster inference on an RTX 3060 by integrating Sage Attention, a third-party optimization node that accelerates attention computation without quality loss. The workflow also adds first-frame and last-frame transition controls and fixes audio sync issues from an earlier version.
LTX 2.3 is Lightricks' open-weight video diffusion model released in late 2024, capable of generating 121-frame clips at 24fps. The base model runs on consumer GPUs but benefits from memory and compute optimizations—Sage Attention replaces standard attention kernels with fused implementations that reduce overhead. The workflow supports GGUF quantized weights, further lowering VRAM requirements for 12GB cards.
Performance gains
Testing on an RTX 3060 12GB showed generation times drop from roughly 4 minutes to under 3 minutes for a 5-second clip at default settings. The updated workflow ships as a JSON file for ComfyUI with nodes for Sage Attention, transition frame conditioning, and audio track alignment. Installation requires the Sage Attention custom node, available on GitHub.
The workflow prioritizes practicality over feature bloat—no upscaling chains, no multi-stage pipelines—keeping it runnable on hardware most users already own. The developer is fielding setup questions in YouTube comments and plans further updates as LTX evolves.