New Animation Update: WAN 2.5

[Viridian]

This one was a bit surprising for me. I thought I had a handle on the main animation apps and their strengths and weaknesses, namely Kling and VEO. Some recent uploads on DeviantArt showed a model I hadn't seen before. When I perused through my OpenArt list, I noticed that WAN had a new update.I hadn't originally been a fan of WAN, but it has been more open and less restricted. It's far more friendly to NSFW content and openly allows characters to go nude, perform sexual acts, etc.

The 2.5 update greatly enhances physics, at least for mud textures, and it can render sound and voice, though not to the same quality as VEO.

Currently, it seems that WAN 2.5 may be the best for rendering specific quicksand textures, easily adhering to the prompts for not only submersions, but also controlled descents (e.g. sinking to shoulders, waist) whereas Kling will just dunk them straight down. Using WAN is less like riding a wild horse.

WAN notably renders visuals with better movement speed. Kling sadly has a strong habit of making its videos render in slow motion that is unavoidable in many renders. On the flip side, WAN struggles more with rendering off-screen items - it has a very hard time cohesively making limbs and hands appear correctly (e.g. characters with their arms buried will pull them up, but the arms are actually too short and bend in wrong directions), and movements can appear too jittery.

I'm doing more experiments with WAN. I might lean towards WAN 2.5 as the best animation model specifically for interaction with mud, but it's far behind Kling for general animation, and VEO for sound. I doubt most people would have subscriptions to multiple models unless you are using an AI creative platform like OpenArt that allows you to flip between them.

[hiimawesome]

Thank you for figuring out what ai gen that one guy was using! It didn't look like anything I'd tried before. I'll give WAN a go.
I mean, it is only one letter shy of WAM

[BogDog]

I saw that at Seaart but the cost makes me hesitate. It is twice as expensive as SORA 2 (1000 points versus 520 for 10 seconds). I can't afford to double my spending.

In the description they praise the audio upgrades too:

"Wan 2.5 is a powerful AI model that transforms your ideas into high-quality videos with perfectly synchronized audio. Moving beyond silent visuals, Wan 2.5 AI Video Model features a built-in audio generation system that can create everything from dialogue and soundscapes to background music, all from a single prompt. It is engineered to understand complex narratives, cinematic styles, and detailed audio instructions, launching a new era of AI-driven content creation.

- Create Longer, More Detailed Stories: With the ability to generate clips up to 10 seconds long—doubling the capacity of previous versions—Wan 2.5 allows for more complete narrative arcs, detailed actions, and greater creative expression in a single generation.

- Bring Your Vision to Life with Precision Control: Thanks to a significant boost in natural language understanding, Wan 2.5 AI Video Model can interpret complex, multi-step prompts. It excels at following commands for continuous changes and specific camera movements (e.g., "zoom in," "pan left") and responds effectively to structured, detailed instructions.

- Experience Stunning Cinematic Quality: The model delivers major improvements in dynamic motion, structural stability, and cinematic aesthetics. It can produce professional, visually appealing videos at up to 1080p resolution and 24 frames per second.

- Animate Your Images with Unmatched Consistency: Transform still images into videos while maintaining excellent character and object consistency. Wan 2.5 provides precise semantic control, ensuring the subject from the source image remains stable and recognizable throughout the video.

- Let Sound Be Your Director: Use any audio file—dialogue, music, or sound effects—as the creative driver for your video. Combine your audio with a text prompt or a starting image to generate a video that is perfectly timed and thematically matched to your sound reference."



Sound is an area where SORA 2 is weak. Some of my vids are rendered with the girl speaking the instruction text in my prompt.

https://www.seaart.ai/models/detail/a70 ... bfbbwan25v

[Viridian]

A montage of my Cheer Squad submergences.

WAN-2.5 EASILY eclipses other AI platforms when it comes to adherence and quality of output. It does have some unique artifacting that it shares with VEO, but it has more of the creativity that Kling is known for. WAN tends to maintain cohesion better, though it has some noticeable visual tics when rendering scenes. The visuals warp less, and even with the visual artifacts from the changing mud texture, it has the best adherence to descriptions of the mud and how the characters interact.

A curious side-effect of the better prompt adherence is that you need to be notably more specific in the prompt. It's less intuitive when pairing expression and action. If you prompt WAN to make a character sink, it will literally do so with little action. You will need to add more detail to show body movements, facial expressions, etc. Kling, VEO and Sora are better are putting together a cohesive output based on context.

[Sandy Place]

These keep getting better and better.
The texture of the quicksand is looking more and more realistic as are the expressions.

Thanks for sharing your efforts and please continue.

[Viridian]

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

[BogDog]

That's looking really good.

[Duncan Edwards]

This one was a bit surprising for me. I thought I had a handle on the main animation apps and their strengths and weaknesses, namely Kling and VEO. Some recent uploads on DeviantArt showed a model I hadn't seen before. When I perused through my OpenArt list, I noticed that WAN had a new update.I hadn't originally been a fan of WAN, but it has been more open and less restricted. It's far more friendly to NSFW content and openly allows characters to go nude, perform sexual acts, etc.

The 2.5 update greatly enhances physics, at least for mud textures, and it can render sound and voice, though not to the same quality as VEO.

Currently, it seems that WAN 2.5 may be the best for rendering specific quicksand textures, easily adhering to the prompts for not only submersions, but also controlled descents (e.g. sinking to shoulders, waist) whereas Kling will just dunk them straight down. Using WAN is less like riding a wild horse.

WAN notably renders visuals with better movement speed. Kling sadly has a strong habit of making its videos render in slow motion that is unavoidable in many renders. On the flip side, WAN struggles more with rendering off-screen items - it has a very hard time cohesively making limbs and hands appear correctly (e.g. characters with their arms buried will pull them up, but the arms are actually too short and bend in wrong directions), and movements can appear too jittery.

I'm doing more experiments with WAN. I might lean towards WAN 2.5 as the best animation model specifically for interaction with mud, but it's far behind Kling for general animation, and VEO for sound. I doubt most people would have subscriptions to multiple models unless you are using an AI creative platform like OpenArt that allows you to flip between them.

Thank you for the detailed review on WAN 2.5. The examples you've posted are really getting there. It's great stuff but most of the feeling I got from my brief experiment left me with the a feeling that the next version of Kling is going to be really awesome if it tops this. The major overriding thing for me is being able to get at least some cheesecake nudity without the dang thing having a stroke. I've spent a lot of time learning to fool Kling into giving me more what I want but it would be nice to be able to rely on dangerous words like "cleavage" not being a problem. You have to keep in mind that Kling is Chinese and the translation does not appear to be 100% spot on. There are moments when you can fool it if you're aware of this and get a little clever with it.

Thanks again and keep going.

[MadMax359]

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

and i love how she's able to get her upper body out & above the surface to keep struggling

[Viridian]

Thank you for the detailed review on WAN 2.5. The examples you've posted are really getting there. It's great stuff but most of the feeling I got from my brief experiment left me with the a feeling that the next version of Kling is going to be really awesome if it tops this.

I've pretty much jumped from Kling to WAN at this stage. Aside from the output style and quality, the usability of WAN makes it so much more worth it.

WAN is markedly more costly to use, even when dropping from 1080p to 720p HD (1080p is twice as costly, and 720p is still more costly than Kling). However, WAN-2.5 feels like it actually has its training modelled on mud/qs scenes, whereas Kling doesn't feel like it has a natural interaction and instead it fumbles through its physics.

Massive bonus points to WAN for actually being able to generate its own realistic mud textures, so you can turn an upper-body scene and make the character sink into a real mud substance. Before I had to use photomanip trickery to create a plausible transition, but I can now create an entire scene just from a single non-QS starting frame with WAN.

Kling clearly isn't trained on it, and the work we do is really trying to gaslight it into thinking that the substance is some kind of sand and not brown water. Half the time Kling turns the scene into water, and when it does keep to the sand/clay theme, it takes it literally and struggles to generate outside details and defaults to clumpy soil texture.

This is evident in that instructing WAN to create struggling actions makes it mimic a style that even looks like what you'd find in an MPV video. You can get arms stuck, shoulders twisting, side to side motions, reaching, and the submersion with the mud correctly closing over their head. It also has vastly superior sound generation compare to Kling, and less buggy than VEO 3.1. The only downside is that its speech generation is mediocre compared to VEO.

I do feel that Kling is dropping off. At least for the QS side, WAN is _the_ gamechanger. Since Ive still got Kling credits, I'll likely use Kling for non-QS scenes. The dynamic movement and camera work for Kling is still superior, though 2.5 really has a bad habit of reverting to floating slo-mo movement. I might throw in headshots of speech using VEO.

But as high praise, I'd label WAN as the first "real" quicksand-trained AI animation model.