Gemma abliterated
Howdy! LOVE the workflows.
Trying to drop in the abliterated gemma models released on Comfy repo yesterday (and have tried a couple of others, too), none seem to work with your workflow or the official. I get an "invalid tokenizer" error as soon as I hit go. One of the other ablits will get as far as as attempting the prompt enhancer before I get a mat mismatch, but I'm not as worried about that one.
Any idea what's going on? Figured if it was on Comfy repo at least THEIR workflow would work, but it's giving the same tokenizer error.
Yes it seems there is something with the dual clip loaded combo.
Kijai is trying to fix that I think : https://huggingface.co/Comfy-Org/ltx-2/discussions/7#69aeab2351b3c6cdaf2d4024
Gotcha. See now it was ACTUALLY intended as a lora rather than a full text encoder replacement, but maybe they'll get it working. I'll just be patient. Can't train my loras to take advantage yet anyway. :D
Ah yes, now i see in your screenshot you loaded it in the clip loader. Thats not correct
Its a lora yes, and for the prompt enhancer node ;-) (not the positive and negative prompt directly)
Which also should have been obvious by the fact that it was only 600MB... it's early, not enough coffee yet... :D
Any tips on adding it to your workflow?
You can put a Lora (model + clip) loader node (its a standard comfyUI node), between the main Clip loader and the clip input at the LTX text generate node (prompt enhancer).
But it seems like it needs a fix from Kijai first.
(I'll add it to the workflows when its working also, since getting a prompt refusal in the prompt enhancer is not ideal ;-) although i dont know how censored Gemma is )
Something like this should probably work eventually (but not yet for this specifically combo) :