Audio ID LoRA and GGUF models
Hey Rune, I'm guessing you have to use the dev model for this? Because I tried with the gguf models and the results were nightmare fuel.
Should work with gguf. Nothing particular about guff, its just a "zipped" format. Kinda...
But that being said, not all gguf are "zipped" the same way, and some might even be a bit incompatible with the split files used here (vae, text encoder etc).
Where did you get the GGUF files from? I have used Quantstack and Unsloth ones myself, without issues
And at what quantitation level? Q2 might be stretching it too far.. but Q3 is decent, and Q4 and up are nice quality
That being said I never tested ID-Lora with gguf, but in theory it shouldnt matter. But will try it out just to check ;-)
🤔 hmmm weird... Visually it looked fine, but the audio was insane. I am using QuantStack LTX-2.3-distilled-Q5_K_M.gguf and unsloth gemma-3-12b-it-Q5_K_M.gguf. I have a real clean audio clip so I toggled off MelBandRoformer.
odd, the vae for audio is not part of the GGUF, but the ID-Lora does "change" the audio, and some attention stuff etc. so maybe its in the render process itself
So maybe its not happy about GGUF, but will try out my end
With LTX-2.3 DEV GGUF Q5 UD from Unsloth + distilled lora (for faster generation) + ID-Lora for the audio
Seemed to work well my end .. strange
(I'll try the distilled gguf too just to check)
Worked fine with distilled too.
Only thing is perhaps the guidance scale at the ID-Lora node. Mine is set to 3 passes (high quality)
With 1 pass it might "skip" sometimes.. but should work most of the time
OK I'll tinker with it some more. Thanks for checking it out, I must have something configured weird.
Try other seed, and other audio file also.. just to rule that out .
On a few rare occasion I have ran into LTX just refusing to cooperate, such as refusing to make the input image come alive.. instead doing a slideshow or refusing to talk instead doing narrative voice over.
But changing the seed, and prompt can sort that out
I tried the v.1.1 version of LTX GGUF as well since that was a plausible scenario where things could break with new improved audio stuff in the LTX Distilled v.1.1 version (and this version coming out after ID-Lora was made). But worked ok too .. .
Yea looks like I just have to roll the dice a bit with the seeds, and some of them turn out ok