Man Beta 6 Turbo is amazing on 4-step with my cooked Lora
finaly got something i was looking for
the beta 6 is working perfect with DMAD and PDMD lota at 4 step
so i merge those to Lora and used on 10eros beta 6 turbo at 4-step got wondefull result
maybe on your next release think about DMAD and PDMD lora for lower step at very high quality
only one issue , not sure you aware of that, the audio reference on ref2v completly ignored!
please share the loras you are talking about, i am curious to test them.
so i merge those to Lora and used on 10eros beta 6 turbo at 4-step got wondefull result
What strength weights and blend, just 1.0+1.0, and what strength is that result used on top well?
Yes i merged them together first and then used as turbo lora on 10Eros
But at the end there is big issue with 10Eros beta 6 by itself audio reference not working correctly which i had to remove the all Adaln to make it wiork
Yes i merged them together first and then used as turbo lora on 10Eros
But at the end there is big issue with 10Eros beta 6 by itself audio reference not working correctly which i had to remove the all Adaln to make it wiork
I've had no audio reference issues. It is way better at audio reference, voice tone reference, and music/sound reference than any of the Eros models, most were built on only less than half of reference ability from the hybrid base. If you mean voice:
subject_definitions:
<Audio 1>: the vocal-timbre reference providing pitch, tone, and delivery style for <Subject 1> (S1) in .
retention_analysis:
<Audio 1>: reference - vocal timbre and delivery guide speech of <Subject 1> (S1) without copying audio signal.
overall_soundscape:
..... <Subject 1> (S1) speaking in the timbre of <Audio 1>.
upload your merged version of lora and even maybe the whole merged model either here on Huggingface or Civitai - please!
The beta 6 is a big leap forward. Fanatstic work. And with the dmad speedup lora it is even working with 4 steps. Fantastic.
I tried the pdmd-lora too. I did get more motion there but at the prize of a totally overcooked result at the end of my 30 second long clip with continuuity.
The quality still is better without those loras and the speed with 10 steps is not slow. Thanks to everybody doing such a good job.
I've tried dmad and some other 4 steps turbo loras. They all make the audio worse and give you this plastic/wax look. I'm running the int8 at 10 steps with spectrum feature forecasting to skip some steps (euler and shift scale 10). I can't see any loss of quality.
SO THIS MODEL BETA 6, is not good for REF2VA?
Actualy is realy good for Ref2VA but audio referece still has issue, i couldnt get the audio referece correctly
I'm beginning to think that he doesn't actually know and he had an llm do it and he's saying that he did it. lol
@stphnvdb Hey Stephan, did you recently update comfyui? Spectrum has quite a bit of new options in their node.