Echo Chamber - AceStep 1.5 song (XL version)
Echo Chamber \(XL version\)
As an experiment I regenerated my Ace Step 1.5 song using XL model (same parameters etc.). It's similar, but there are differences. I've noticed that the old 1.5 would sometimes improvise a bit to fit lyrics better to the song, while XL will more often rush with lyrics and leave a pause. I've had yet another version of this song, that failed to generate properly with 1.5 (with interesting results), but would properly generate using XL model.
I'm not sure I like the XL version of this song better, but XL tends to be better with following lyrics (if somewhat less flexible).
Here is the non-XL version of this song (with prompt, lyrics, etc.): https://www.reddit.com/r/AceStep/comments/1sf99em/echo\_chamber\_acestep\_15\_song/
I've also noticed that the text encoder for Ace Step isn't 100% deterministic. Haven't boiled down which factor is causing this, but if I run AceStep with same parameters (seed, model. prompt, the whole shebang) on a different machine, I'll get a different song. I still get the same song on the same machine though. It might be tied to OS, pytorch or ROCm version (not sure which). Previously I thought it was a change in ComfyUI (that might have been true at some point in the past), but I was wrong (otherwise I wouldn't be able to generate this version of the song).
https://redd.it/1sikd31
@rStableDiffusion
Echo Chamber \(XL version\)
As an experiment I regenerated my Ace Step 1.5 song using XL model (same parameters etc.). It's similar, but there are differences. I've noticed that the old 1.5 would sometimes improvise a bit to fit lyrics better to the song, while XL will more often rush with lyrics and leave a pause. I've had yet another version of this song, that failed to generate properly with 1.5 (with interesting results), but would properly generate using XL model.
I'm not sure I like the XL version of this song better, but XL tends to be better with following lyrics (if somewhat less flexible).
Here is the non-XL version of this song (with prompt, lyrics, etc.): https://www.reddit.com/r/AceStep/comments/1sf99em/echo\_chamber\_acestep\_15\_song/
I've also noticed that the text encoder for Ace Step isn't 100% deterministic. Haven't boiled down which factor is causing this, but if I run AceStep with same parameters (seed, model. prompt, the whole shebang) on a different machine, I'll get a different song. I still get the same song on the same machine though. It might be tied to OS, pytorch or ROCm version (not sure which). Previously I thought it was a change in ComfyUI (that might have been true at some point in the past), but I was wrong (otherwise I wouldn't be able to generate this version of the song).
https://redd.it/1sikd31
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community
Built a local browser to organize my output folder chaos -- search by prompt, checkpoint, LoRA, node type, etc
https://redd.it/1siqf2v
@rStableDiffusion
https://redd.it/1siqf2v
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit: Built a local browser to organize my output folder chaos -- search by prompt, checkpoint…
Explore this post and more from the StableDiffusion community
Why is Wan 2.2 N.S.F.W Remix Lightning Model so much better at things like hair flip, hair combing and feminine energy than regular Wan?
I am not talking about actual N.S.F.W I am talking about the model that has such a name in it, and just feminine energy, seductive performance, shampoo commercial hair toss, sensual movements, elegant leg cross sitting on bar stool.
Whenever I use any of these WAN models it comes out very static and it ignores the prompt, when I use the remix it comes out nearly perfect.
It's almost like using Grok, not the new Grok but the old one before it was censored.
https://redd.it/1sipeko
@rStableDiffusion
I am not talking about actual N.S.F.W I am talking about the model that has such a name in it, and just feminine energy, seductive performance, shampoo commercial hair toss, sensual movements, elegant leg cross sitting on bar stool.
Whenever I use any of these WAN models it comes out very static and it ignores the prompt, when I use the remix it comes out nearly perfect.
It's almost like using Grok, not the new Grok but the old one before it was censored.
https://redd.it/1sipeko
@rStableDiffusion
Reddit
From the StableDiffusion community on Reddit
Explore this post and more from the StableDiffusion community