Player not loading? Watch on YouTube
This tutorial installs LTX 2.5 through Wan2GP inside Pinokio for local AI video generation with synchronized audio. The speaker updates Wan2GP first and reports version 1250 as the version used. Starting a generation triggers automatic downloads of the selected checkpoint and supporting files.
The comparison covers the full 22-billion-parameter Dev model, Distilled, and Distilled NVFP4. The speaker describes both distilled variants as eight-step models and chooses them for an RTX 3080 with 12GB VRAM. Reported downloads include roughly 19.5GB for the Distilled checkpoint and around 23GB for its complete installation. NVFP4's main checkpoint is reported at about 14.4GB, with additional files required.
After loading the model, NVFP4 generates a 10-second horizontal clip at 720p in about three minutes. Distilled finishes roughly ten seconds faster in the speaker's repeat test. Initial runs take longer because they include downloads or model loading. The comparison with earlier LTX 2.3 results uses different resolutions, so it does not establish a controlled speed benchmark.
A text-only Goku prompt fails to produce the intended character. A reference image produces motion, but the face changes between frames and the result looks three dimensional. The speaker reports a local resolution ceiling of 1080p and says the media flow filters from an earlier workflow are not yet compatible.
