Gemma lora
Curiosity got to me, but not quite sure how to use the Gemma lora.
Any hints? ;-)
Since its a "text model" and all ...
First time around for a "text lora" so i was not sure ;-)
(or maybe it something to come .. . for later)
Abliteration is meant to remove the refusal from the model (which Gemma3 is bit too prone to fall into) when generating text.
So the use case for this is the prompt enhancer for example, generally it should not be useful or can even be detrimental when used for prompt encoding though, so it makes sense as a LoRA you only apply to the prompt enhancement "clip".
yes figured it would be for a bit more "liberated" mode.
Prompt enhancer makes sense ;-) understood now
I tried doing as the picture posted above but there seem to be no difference. I wrote a tame prompt then increased it until the prompt enhancer with the normal text encoder refused. Then loaded the lora to test and got the same refusal.
I tried doing as the picture posted above but there seem to be no difference. I wrote a tame prompt then increased it until the prompt enhancer with the normal text encoder refused. Then loaded the lora to test and got the same refusal.
That's odd, I suppose it's not 100% certain to work? I also uploaded the "heretic" version which people say is better, but didn't test it much. I also tested with pretty tame prompts and compared to the full ablit, and didn't get any direct refusals, for example:
I tried doing as the picture posted above but there seem to be no difference.
I removed the pictured so it would not confuse anyone looking at the thread.
But basically same as picture, but connect it to the LTX prompt enhancer node (and not to the positive negative clip inputs)
But maybe thats already how you connected.
EDIT:
Ah i see Kijai replied at the same time, with screenshot and all ;-)
That's odd, I suppose it's not 100% certain to work? I also uploaded the "heretic" version which people say is better, but didn't test it much. I also tested with pretty tame prompts and compared to the full ablit, and didn't get any direct refusals, for example:
Yes that works! My problem was that I loaded the Gemma 3 text encoder with the DualCLIPLoader node. (Tried 3 different versions with text projection or ltx-2.3 model as second clip) And apparently that breaks it. But just doing the Load CLip node as in your picture works, thank you!
That's odd, I suppose it's not 100% certain to work? I also uploaded the "heretic" version which people say is better, but didn't test it much. I also tested with pretty tame prompts and compared to the full ablit, and didn't get any direct refusals, for example:
Yes that works! My problem was that I loaded the Gemma 3 text encoder with the DualCLIPLoader node. (Tried 3 different versions with text projection or ltx-2.3 model as second clip) And apparently that breaks it. But just doing the Load CLip node as in your picture works, thank you!
Right, yeah this is true, the keys won't match for the wrapped LTXAVTEModel... will have to fix this, thanks for the report.