I keep having fun with ref2va model.
RTX 3060, 64 Gb RAM. I use ref2v Turbo 4 step Lora paired with Sol Attention and Minimax H3 Memory Effecient Sage Attention at 6 steps. It takes about 2 minutes per second of generation.
I've been using LTX 2.3 for quite some time but as soon as I wanted to make just a few simple shots of the same character with cuts, LTX wasn't even remotely capable of that. Which left me so frustrated I eventually gave up on it completely.
But when I tried Minimax everything has changed. Reference to video model is something else. Honestly feels like magic. Being able to put any character into any environment with any custom audio is just mind-blowing compared to what the open-source community had before.
So now instead of constant frustration, I feel pure joy and excitement about the results.
It takes about 10 min per 5-second clip with my RTX 3060 and 64 Gb RAM. The latest shots were even easier to control because of the new KJ preview node.
I've had this idea for quite some time. But only with the brilliance of Krea 2 it has become possible. Truly an amazing model.
Cast:
Sean Connery as Neo
Audrey Hepburn as Trinity
Michael Caine as Morpheus
Clint Eastwood as Agent Smith
Steve McQueen as Agent Jones
Alain Delon as Agent Brown
Marilyn Monroe as The Woman in the Red Dress
I've been practicing some clothing/physics. After trying to fit the downloaded assets on the metahuman I figured out it would be way easier to do in Marvelous, which it was.
I've also found out you can just transfer the weights from any skeletal mesh directly in Unreal. That's what I did with the shoes. I used to do skinning in Blender but fixing and painting the weights was a nightmare. Now I can do the transfer in a couple of minutes.
P.S. The top is yet to come.