Kijai

/

LTXV2_comfy

like 353

If using LTX-2 dev as the main model, you can use a distilled lora to run as if it was a distilled model (cfg 1, 8 ish steps).
The resized_dynamic are alternatives to the stock default lora for same purpose. The stock default one is 7GB, and the resized dynamic far far less, but seems to do the job as good, and easier on the system requirements

Great! thanks for your answer.

Indeed it's the distill LoRA with dynamically reduced ranks per layer, retaining most of it's function while reducing size.

Then, may I ask, if I use the main model from the distilled version instead of the dev version, would I no longer need to use this LORA?

Then, may I ask, if I use the main model from the distilled version instead of the dev version, would I no longer need to use this LORA?

yes exactly ;-) the lora is only needed if you want to use DEV model with low steps (as if it was distilled model)

Then, may I ask, if I use the main model from the distilled version instead of the dev version, would I no longer need to use this LORA?

yes exactly ;-) the lora is only needed if you want to use DEV model with low steps (as if it was distilled model)

thanks for your answer，I get it！✅