Mega v6 still has problems with keeping the same face.
Any hints on how to solve this, or should I just ditch the entire Mega i2v line and go back to v10?
I'm using 'reference image' with 'trimvideolatent' (even just 1 frame is enough most of the time) which helps abit + reactor, but reactor sometimes do weird face placement when the face is partialy covered (like when the person is turning around for example) ... but still, experimenting with these two helps
Yes, it is better then first releases but it changes the face. And also very weird thing is that it "washes" colors on the first frame(s) of initial image (I2V).
I did some minor tests yesterday and Mega v6 is much better in keeping the character consistent. The movement is much better too. The only way to keep your character consistent as much as possible in longer videos is to train your own loras.
I will upload a new version of my 4.0 Workflow where I did some changes. When the prompt has an end frame I will not take the last x images from the last batch anymore but just the one image as new start frame for the next batch. At that point the character is 100% consistent again. The drawback is that the final video can look "interrupted in movement" or "freezed" at this point cause the movement is interrupted.
I'm more talking about i2v, not t2v...
Reference face: frontal view of the woman with defined features
Identity lock: true
Reference face: frontal view of the woman with defined features
Identity lock: true