Request for a Non-Turbo Beta 6/7 – Better Physics vs. Visual Quality Trade-off
Hi TenStrip,
I've noticed that Beta 6 has a trade-off between aesthetics and physics. With fewer steps (8), the visuals look better, but the physics and prompt following are worse. With more steps (10–12+), the physics improve, but the characters look noticeably less appealing.
Beta 5 Non-Turbo at 20–30 steps gave me much better overall results, especially in terms of aesthetics, motion accuracy, and prompt following.
Would you consider releasing a Non-Turbo / Non-Fast version for Beta 7?
Personally, I don't care about generation speed. I'd much rather wait longer for the best possible quality. I feel like Turbo/Fast approaches sacrifice too much quality, and I'd strongly prefer a version without them entirely.
Thanks for your work!
Honestly I don't really understand what the point of making the thing exclusively turbo is.
Like why not just add or link a turbo lora in the repo?
Honestly I don't really understand what the point of making the thing exclusively turbo is.
Like why not just add or link a turbo lora in the repo?
As far as I have been told loras and the patched parts of the model are stored in system memory, so if you apply a lora it increases the duration of the generation, so if true there is a reason for having a turbo model considering the whole point is speed. At any rate even if it's not true I have noticed when you throw a bunch of loras at a model it slows it down significantly so... even if the mechanism isn't true there is a slow down from loras. Also I noticed worse results from beta 5 non-turbo in every respect so...
Also I noticed worse results from beta 5 non-turbo in every respect so...
Which loras did you try using?
Also I noticed worse results from beta 5 non-turbo in every respect so...
Which loras did you try using?
including no loras, well specifically my issue was an anatomy issue and the whole scene falling apart, where as with the same prompt I would get descent results with the turbo model, including going to 30 steps without turbo. Although as far as I remember beta 4 non tubro I think worked fine so I'm gonna assume whatever broke beta 5 non-tubro is still in the model hence why no beta 6 non-turbo, and I'm obviously not talking about minimax in general, for the base model and DesiwaMinimaxH3 hybridV2 no turbo does deliver better results, even just general visual quality, never mind physics. but eros10 beta 6 looks better and sounds better when doing photo realism than the base model or desiwaminimaxH3 V2 w/30 steps so... Yeh beta6 is pretty awesome apart from it not following prompts as well as beta 5 turbo.\
Will admit that I'm going for some extreme / unrealistic body types so maybe I'm being too ham fisted for the base model and the terms I am using are breaking the anatomy "<Subject 1> (S1), the Chinese dwarf woman from the character reference sheet in <Picture 1>, athletic extreme hourglass figure, pendulous saggy br3asts, skinny narrow waist, and wide hips with hip dips, fat BBL a55, " for example and yeh obviously I'm not using the numbers for words in the actual prompt I just don't know if the discussions here are censored.
And just to be completely clear because it could look like I'm trying to generate something I'm not, the tag "woman" will leads to adult females who look 30ish, so no the other thing is not the goal... seriously try it clothed and in a non-sexual context yourself if you don't believe me with either beta 5 or beta 6, I think if my memory serves correctly beta 4 and even the base fl2va / ref2va all react the same way, woman is a strong anchor to the average age for a female, aged down somewhat to about ~30 considering that the bias of the training will be for prettier woman just due to just the available dataset over representing prettier women form 18 - 35, so it averages out at about 30, even if you go about trying to culling the dataset a reasonable amount.
The sole goal is a size difference, not the other difference.