Phr00t

/

WAN2.2-14B-Rapid-AllInOne

like 1.14k

I'd say that is too much motion (multiple jumps) to expect from this Rapid AIO in a single prompt. Frankly, I've kinda given up on expecting lots of motion and prompt adherence straight from WAN alone. It is possible the full WAN with "high" and "low" models, without accelerators, would do better here. However, I've found the real solution is combining WAN with Qwen Image Edit. I can get just what I want, so much easier and faster (because I can see the start and end frames much faster than waiting for any WAN generation).

For example, you generate the first frame with my Qwen Edit AIO:

Then the final frame, you "edit" and tell the fox to be on the other side:

Then it is a very simply WAN first to last frame (with Mega v7):

I think you're right about the motion prompting - this is something inherent in all WAN models, not just RapidAIO.
Which is a shame as it feels like a lot of the hype promised around WAN 2.2 has failed to materialise... Either that or you just need 100GB VRAM to take advantage of it

Maybe one stupid question: the help of Mystic XXX Lora 's - Lora used is I2V or ?

Maybe one stupid question: the help of Mystic XXX Lora 's - Lora used is I2V or ?

It seems to really work well on T2V as well, despite being a. I2V LoRA! Though like with most NSFW LoRAs, it is very prompt dependent. Sometimes it works brilliantly and sometimes... Yeah, not so much 😂

I will note that the NSFW Mega v7 does include some of MysticXXX already.

I will note that the NSFW Mega v7 does include some of MysticXXX already.

Some? What does that mean? Are there multiple MysticXXX LORAs (I'm not familiar with Mystic)?

I will note that the NSFW Mega v7 does include some of MysticXXX already.

Some? What does that mean? Are there multiple MysticXXX LORAs (I'm not familiar with Mystic)?

Meaning it isn't added at 1.0 strength, but applied as percentage of the whole mix.

I really do not know am I the only one here that has pale videos in FLF (mega AIO) ? Motion ok, prompt adherance ok, but colors no. I am doing everything according to recomendations .... :(

@Seeker36087 I edited the prompt slightly. He didn't do any flips, but he jumped across 2 boulders.

Mega v7 NSFW, shift 8, sa_solver/beta

"Super fast motion. Cinematic movie, modern Animation style, Japanese ink style influences. Smooth cinematic motion and pace, dynamic camera.

A fast-flowing river in the centre of the frame with riverbanks to the left and right of the frame. Six large mossy boulders span the distance of the river between the two riverbanks.

An anthromorphic fox ninja stands on two feet on the left riverbank. Fox ninja performs a series of intricate leaps, backflips and mid-air pirouettes as fox ninja progresses across the six mossy boulders to cross to the right riverbank. Fox ninja lands from the jump."

I really do not know am I the only one here that has pale videos in FLF (mega AIO) ? Motion ok, prompt adherance ok, but colors no. I am doing everything according to recomendations .... :(

Make sure you are using the mask frames to skip the grey frames (this is taken care of in my example workflow).

Yes, thanks, that was my fault (mask frames were off).

Just trying out the new v8 - the quality of the video seems improved a bit (though there's still an odd fizziness to euler_a/beta) but the slow motion effect is even more pronounced for some reason.
Even when specifically prompting 'smooth natural motion with no speed up or slow motion effect', the resulting video feels like it's playing at around 4fps despite being set to 16fps.
I assume this is the Lightning Loras at work again... I've been trying out RCM myself and have noticed that motion is usually improved but the end quality is almost always extremely blurry and over-smoothed.

Just trying out the new v8 - the quality of the video seems improved a bit (though there's still an odd fizziness to euler_a/beta) but the slow motion effect is even more pronounced for some reason.
Even when specifically prompting 'smooth natural motion with no speed up or slow motion effect', the resulting video feels like it's playing at around 4fps despite being set to 16fps.
I assume this is the Lightning Loras at work again... I've been trying out RCM myself and have noticed that motion is usually improved but the end quality is almost always extremely blurry and over-smoothed.

I only include ~0.3 of WAN 2.2 Lightning in v8, so it is unlikely to be the cause. I felt adding rCM 720p did improve motion, but it remains very prompt and seed dependent in my tests. Using first to last frames really does help take the guesswork out of it.

I'm working on a new Mega model, and I think getting rid of SkyReels completely is going to greatly help motion. In my test building environment, this is what I get for the original fox jumping prompt now:

Here is another random test of a stunt car:

All still 1 CFG and 4 steps and 1 model. I'm working on packing up a Mega v9 which will hopefully significantly improve motion.

For some reason, neither of those videos are loading for me :(

Not sure why those videos are not working... codec issue or something...?

v9 is posted though! Give it a try!

Just been trying the NSFW v9 and I think removing Skyreels was a smart move - motion is definitely better, if still a bit slow motion (but that's the fault of the Lightning loras, not the Rapid AIO model).

One interesting thing I've gleaned from my testing... Setting the image to video node's strength to 0 for T2V as advised produces a pretty good result with slight issues with prompt adherence. But when I experimented setting it to 1 instead while still doing T2V, the adherence seemed to improve! I'll continue testing to see whether T2V is actually consistently better with the strength set to 1... Very interesting!

Excellent work Phr00t. We all thank you for your tireless work in refining this :)