How to use checkpoints

[dropmasta]

The documentation says that these checkpoints can't be used directly for inference. Do they have to be converted to be useable in for expample ComfyUI?

[PGCRT]

Epoch folders contains the ".safetensors" that you can use for inference,
Global_step folders contains deepspeed data in ".pt" format and can't be used for inference

[dropmasta]

Thanks for the clarification! I get an "lora key not loaded" error when using the LoraLoaderModel with the LTX model.

[tdrussell]

I'm not sure if ComfyUI supports loras for LTX Video. I'm guessing not, considering it's not loading the weights. It also depends on whether you're using the LTXV extension, or native ComfyUI implementation. Lora support would needed to be added to both. Probably it's easier with the native implementation, since theoretically all you need to do is support the state_dict key format when loading the lora.

[dropmasta]

There was a lora on citivai that was deleted. It was working fine with the native LTX implementation and LoraLoaderModel.

[jashwanth0712]

I was trying to train a lora of my video , when i am using "Man" as keyword i am able to generate my face in the video , is there any way of keeping a keyword for any character

[dropmasta]

When training image models is use a rare token like "ohwx man" as caption and keyword.

[jashwanth0712]

When training image models is use a rare token like "ohwx man" as caption and keyword.

I could not understand from the comment mentioned in the example , on how to use the caption, shouldit be the filename of that corresponding image , or do we maintain a json for it