fp8 SafeTensor

FP8 version is a Transformer-Only prune at mixed precision. Run it with a diffusion loader like kijai's transformer only versions. Do not change the model weight to fp8, leave it default. If you force load fp8 on the loader you will cast the bf16 audio to fp8 and ruin the audio. The fp8 version is about space-saving not precision. The fp8 may also struggle more with adding audio to video inputs, and using audio inputs for a video.

BF16 is the max quality full checkpoint. If you can load and run the ltx22b dev model, load and run that one instead for the best results.

Conditioning is very important for i2v. If you are doing a simple animation with no pose changes, leave i2v conditioning at 1.0. If you are doing storytelling, camera movement, or complex prompts you must turn down i2v conditioning to 0.6-0.75 to allow the first pass freedom. If you see any kind of body contortion, upper body horror, etc, it is usually because input image conditioning is interfering with the model trying to enact the prompt.