New Animation Update: WAN 2.5 (and WAN 2.6)

[BogDog]

I have noticed some odd artifacts here that I also see in Sora 2 and have been a pain for me. Surface textures change so much over time passed. Note at the end of the video the surface of the mud smooths out and starts to accumulate little black lines.

sample1.png

Seen that even worse with Sora. A sandy gritty texture evolved until it is smooth with crosshatched lines and such, as you can see here with before/after snapshots. Top/beginning shot has a nice sandy look and by the end of the clip the quicksand looks like a comic book drawing. Quite annoying.

texture.changed.png

[Viridian]

Oof, yeah, the artefacting is nasty. I'm finding that I need to use AI to retouch the frames and use editing tricks to transition to new starting frames. WAN has a serious problem losing visual cohesion. All AI animation does, but this is probably the worst I've seen, though it strongly depends on how much movement there is. The more "new" visuals it has to generate, the more it fills it with the pencil crosshatching.

[Acidtester]

That safari girl scene is really incredible!!
The more I’ve watched it the more I absolutely love it. She looks awesome, and the way she kind of jerks around in it really gives the feeling that she’s stuck in some soupy goopy mess that she can’t get out of. It’s so wild!!
I can’t thank you enough for bringing this sort of thing to life. I don’t think it could get any better.
Amazing work!!

[Viridian]

Now for some big artistic development. First step is to take my own original artwork, turn that into a photorealistic AI render, then turn that into an animated scene with WAN-2.5.

[MadMax359]

great progress getting the program to show more skin

[mjw]

Thanks for this research, it's been very useful.

I find with Kling, having the right start image is really important for it to do a decent job on mud textures (and 2.5 is much better than 2.1). If the start image even hints there's water about, you end up with brown water.

I have a flux.io subscription that has multiple models available (and also a Kling subscription, it it weren't already paid up for the year I'd be tempted to drop it right now, on the strength of WAN 2.5). What I typically find best for mud videos is to make a starting image using Nano Banana - it's really good at taking multiple images and making one final image, so typically I will use a photo I took of some real mud, and add an AI generated character I like, or use an image of myself, and get the initial mud scene from that. (I've been able to recreate a number of mud sessions I had, I photographed very few) then use that. Nano Banana isn't particularly NSFW friendly. It will go as far as making images of a couple kissing, but try any nudity and it fails (so it's a good job my preference is for a fully clothed sink). The flux.io interface allows you to upload multiple images and compose a single result.

Seedream 4.0 will do NSFW images, but has a number of drawbacks (it usually takes a few attempts not to have a random hand in the wrong place, or someone with 7 fingers) and it absolutely won't do men with shaven legs.

flux.io also has the WAN 2.5 model, and it's so much better for sinking videos I've upped my flux.io subscription. Although it's more expensive, it's so much more likely to give a decent result on the first try, the total cost is less for me. I think the only real miss I've had with it is one video where the subject suddenly and inexplicably (and about half the speed of light) made a 180 degree turn. But things I've liked about WAN 2.5 is it will correctly have the subject turn their face up when the mud reaches their chin (I've rarely succeeded with Kling, it always has the subject do it when they are nowhere near to neck deep) and all kinds of different mud seems to behave pretty realistically. Although sometimes it will have breathing noises still when the subject has pretty much gone completely under.

Kling now allows end frames including for 2.5 Turbo, which can also really help get the video you want (e.g. prevent the person sinking straight to the core of the earth when you just want them to sink shoulder deep - make the start/end frames with Nano Banana, then use these with Kling).

[Das]

Trying to work in a 1-minute bikini quicksand sequence. This is beginning to look like an authentic amateur shoot.

Excellent job! Mind you, I always add bubbles on the surface at the end when I use Sora and Grok

[hiimawesome]

Thanks for this research, it's been very useful.

I find with Kling, having the right start image is really important for it to do a decent job on mud textures (and 2.5 is much better than 2.1). If the start image even hints there's water about, you end up with brown water.

I have a flux.io subscription that has multiple models available (and also a Kling subscription, it it weren't already paid up for the year I'd be tempted to drop it right now, on the strength of WAN 2.5). What I typically find best for mud videos is to make a starting image using Nano Banana - it's really good at taking multiple images and making one final image, so typically I will use a photo I took of some real mud, and add an AI generated character I like, or use an image of myself, and get the initial mud scene from that. (I've been able to recreate a number of mud sessions I had, I photographed very few) then use that. Nano Banana isn't particularly NSFW friendly. It will go as far as making images of a couple kissing, but try any nudity and it fails (so it's a good job my preference is for a fully clothed sink). The flux.io interface allows you to upload multiple images and compose a single result.

Seedream 4.0 will do NSFW images, but has a number of drawbacks (it usually takes a few attempts not to have a random hand in the wrong place, or someone with 7 fingers) and it absolutely won't do men with shaven legs.

flux.io also has the WAN 2.5 model, and it's so much better for sinking videos I've upped my flux.io subscription. Although it's more expensive, it's so much more likely to give a decent result on the first try, the total cost is less for me. I think the only real miss I've had with it is one video where the subject suddenly and inexplicably (and about half the speed of light) made a 180 degree turn. But things I've liked about WAN 2.5 is it will correctly have the subject turn their face up when the mud reaches their chin (I've rarely succeeded with Kling, it always has the subject do it when they are nowhere near to neck deep) and all kinds of different mud seems to behave pretty realistically. Although sometimes it will have breathing noises still when the subject has pretty much gone completely under.

Kling now allows end frames including for 2.5 Turbo, which can also really help get the video you want (e.g. prevent the person sinking straight to the core of the earth when you just want them to sink shoulder deep - make the start/end frames with Nano Banana, then use these with Kling).

Does flux block NSFW? I can't run any of this stuff locally, so I was curious what services you all used to do this stuff. It seems expensive, like 50 cents for 5 seconds at 720p.

[Viridian]

And just like that, WAN-2.6 is out.

WAN-2.5 was really a "Preview" model for the 2.6 release. 2.6 boasts lots of improvements, notably better cohesion and visual fidelity, better audio sync, voice lines and lip sync. It also can render up to 15s, which is perhaps the longest AI video generation at the moment.

As far as these claims go, they're pretty solid. It definitely has better voice now, which puts it competitively against VEO 3.1 and above Kling 2.6. This makes it decent for storyboards and dialogue.

However, the animation quality is... kind of shit. Like a step backwards. It definitely struggles with the quicksand/mud texture, but it generally doesn't seem responsive to movement prompts.

I ran the same prompts for WAN 2.5 and WAN 2.5 - twisting movements, grunting noises, camera movement. If I didn't label the files, you might have the versions reversed. WAN 2.5 is better in almost every way. Specifically for QS scenes, WAN 2.5 is clearly using models trained on real interaction with mud. 2.6 seems to have no clue what mud is.

The cost is the same as 2.5. Unless you're working with voice lines, I would not recommend WAN 2.6. I'm currently working through a film voiced with Kling 2.6, but I might switch it up with WAN 2.6 to see if it's better.

Edit: No, WAN 2.6 voices are robotic, as is its movements and animation. 2.6 is a huge step down. The audio and voices are better than 2.5, but that's a low bar. It's quality is the lowest against the big models (VEO, Sora and even Kling), and the animation is inferior to WAN 2.5.