Mega I2V problem
Hi there! New Mega AIO is almost perfect, but yes, in I2V mode (despite bypassing end frame image according to creator's instructions) changes some things immediately after the generated video starts. As an old techical person I have tried many "tweaks" but yes, nothing changes that Mega I2V behavior. Mega AIO version 10 does not do that. Change of initial face from the image is the biggest problem. It does not happen with AI generated faces. That was a remark from me, but anyway it is a great work!
Yes, same problem here with I2V. For example, compared to v10 Mega decides to change hair color for no reason. Try to fix that by specifying hair color in the prompt, then it completely changes hairstyle and facial features so it looks nothing like the input image.
Basically I find it unusable for I2V like this. Which is a shame as the output video quality is amazing.
Not an expert but perhaps this is related to the problems people are having with using Loras in Mega. There is something in Mega that is just doing it's own thing and overriding input image and/or Lora?
Since is a mixed t2v and i2v and has accellerators it cause this issues, try adding low noise as lora -0.3
Yes I agree, it has a perfect output but the input face and the output face is completely different (and also an image ambient). For such a cases I will use V10.
just change the strength in WanWaceToVideo to 1.5,should be better
just change the strength in WanWaceToVideo to 1.5,should be better
I'm working on a fix. It is the FunReward LORA that is included causing problems.
Reporting back: great output quality (thanks!) but faces does not look like the input image.
Give "mega" v2 a try. I was seeing much better face consistency without the FunReward LORA. Also prompting some details about the face seems to help.
I tried mega v2, it's a bit better with facial features but it still changes things from the input image too much compared to v10. For other uses it's probably great, I do like the output quality very much, but for my purposes it's just way too frustrating. I'll stick with v10 for I2V.
It is interesting. Mega V2 non NSFW has a much lower changes of an initial photo, but NSWF version does it dramatically from the very beginning. When some action begins, you have totaly another person in a generated video. But ok, the author does his best and for some time I will also stick with V10. The only thing I always have problems with is ( I2V) a camera movement. It simply does not "listen" to a prompt (stays still in a most of a cases).
Yea, I should have specified that I'm using the NSFW versions of both mega and v10. I haven't tried the non NSFW version.
Yea, I should have specified that I'm using the NSFW versions of both mega and v10. I haven't tried the non NSFW version.
I might be able to pull back some of the NSFW LORA strengths to see if that helps. However, I can only pull it back so much before it loses too much NSFW details... I am working on an NSFW "mega" v3.
You might have better luck using "Mega v2" with just a few hand picked NSFW LORAs.
Anyone got a useful prompt to prevent that fluids coming out of the mouth?
Anyone got a useful prompt to prevent that fluids coming out of the mouth?
the vagina opening is the only source of liquid, there is no liquid in her mouth, there is no liquid in her belly button
I've had it literally shoot of of nose ears mouth belly button, lol