Kijai

/

LTX2.3_comfy

like 399

T2V:

https://www.bilibili.com/video/BV1jow2zmEUA/?spm_id_from=333.1387.homepage.video_card.click&vd_source=fc45d39e8b76813ac6bb03d5e41c4de8

Model Name: ltx-2.3-22b-distilled_transformer_only_mxfp8-block32.safetensors
10s duration, hot start, secondary sampling (euler_ancestralcfg_pp, euler_cfg_pp), MXFP8 distillation model with 8 steps:

• 1024 * 600：169s
• 1280 * 736：169s
• 1532 * 832：253s
• 1920 * 1088：483s

In addition, I found that using the character Lora (ltx-2.0) trained by myself would slow down the sampling speed and increase the overall running time by about 100 seconds+

with chracter lora

• 1532 * 832：350s
  without chracter lora
• 1532 * 832：253s

The live action scene feels slightly inferior, but the Pixar animation effect is good and the difference cannot be seen