Seedance 2.0 release

[Viridian]

This is subjective, but I don't see how Seedance has better mud physics. The Seedance scene looks relatively awful. That's a character flailing with their arms and legs in a water puddle.

At the same time though, I don't think you're giving either one a fair shot. The starting image _is_ a puddle of water. WAN 2.5 can turn that into thick mud (I ran that for science). WAN, as we know, has very high prompt adherence - just very poor dynamic movement and - especially by current standards - poor visual quality.

wan.mp4

If you actually use a source image that has a decent mud texture, the results are quite different for Kling.

kling3.mp4

It can't be understated how important it is to use the right image reference. Most creators who are struggling with getting good results haven't figured out what Kling responds well to - or how to set up scenes so that Kling can use them. One of the hidden abilities is that Kling can in fact turn "anything" into mud/QS like WAN can. In this scene, my character walks over flat sand, which turns into mud underfoot, which then sets up a sinking sequence.

kling_sand.mp4

kling_sand2.mp4

[Sekani]

At the same time though, I don't think you're giving either one a fair shot. The starting image _is_ a puddle of water. WAN 2.5 can turn that into thick mud (I ran that for science). WAN, as we know, has very high prompt adherence - just very poor dynamic movement and - especially by current standards - poor visual quality.

I know it's possible to get decent mud with Kling (your portfolio is evidence enough of that, and I was managing myself back on Kling 1.6), but I figured a more "fair" comparison would be with both models just out of the box. The source image was pretty low effort I'll admit.

If anything, if that Seedance clip is all I'm getting for 50 cents per second, that's a significant point in Kling's favor.

[Viridian]

...but I figured a more "fair" comparison would be with both models just out of the box. The source image was pretty low effort I'll admit.

I understand this approach. We've clearly seen that neither Kling nor Seedance can generate a mud/qs texture on its own, and neither can turn a wet texture into thick mud/qs. However, there seems to be assumptions about output quality based on the "worst' inputs, so I was more interested in what Seedance 2 could do with a good input, given that we have excellent exemplars of Kling 3.0 as a benchmark.

My hunch during this time between announcement and launch is that Seedance 2 is overhyped. The marketing heavily pushes action sequences, but Seedance appears to be mediocre at everything else, while Kling is pretty much as good at dynamic scenes at a fraction of the cost. Seedance may indeed be better with action scenes involving interaction with other characters and environment -- laughably, I can barely get Kling to make a character swim.

Ultimately, if Seedance 2 can't render QS at all even with good inputs, then the cost doesn't matter.

[Green Heart]

The problem with AI models is you have to figure out how to make them dance. I figured out how to make Sora 2 generate really good looking quicksand videos and images, but a lot of people didn't. I didnt make my methods or results public because I was curious at if i was lucky or not. I kinda wanna see if I can figure SD 2.0 out...

[Sekani]

I don't know if I hit the lottery on this one or what, but the text-to-video prompt was something along the lines of "a buxom jungle girl crawling through the mud before sinking into it" and this what Seedance 2.0 gave me. I was shocked.

[Sekani]

Same source image, same basic-ass prompt: "the woman is stuck and sinking in quicksand"

And yes this an anime image, I was in the mood for something different.

Wan 2.5

Wan-2.5-Preview-a82b0928-96f1-494b-a537-3e6ca9aac05a.mp4

Kling 3.0

Kling-V3-Standard-f7ef8ab1-0cc1-4e9b-a6c7-6b6711069283.mp4

Seedance 2.0

Seedance-2.0-f8a3d966-4d0f-4226-967b-2da3e30491e4.mp4

[Sekani]

I don't know if I hit the lottery on this one or what, but the text-to-video prompt was something along the lines of "a buxom jungle girl crawling through the mud before sinking into it" and this what Seedance 2.0 gave me. I was shocked.

I wanted to see if this was a fluke, so I ran this as a text prompt: "A buxom jungle girl wearing a tattered loincloth is stuck in thick, heavy mud. Her struggle to escape the mud causes her to sink further into it." All I can say is that this is not the same model I tried out a week ago.

[Viridian]

Seedance 2.0 is now released on AI art sites. i gave it a test run using a thick dry mud texture. I ran the same prompt with Kling 3.0 - same resolution (720p), same length (10s -- though both are capable of 15s, but Seedance is crazy expensive).

The prompt was:

woman steadily sinks deeper into thick dry mud, she looks down nervously, exerted grunts, she sinks under the thick mud

Test Summary:
- Seedance 2.0 fails the test. Produces water even with a dry visual and text prompt. Quality is pretty good, but for its price, absolutely not. IMO, can generate decent outputs from text to prompt, but overrated.

- Kling 3.0 is IMO the industry standard. It has some annoying sound artifacts and can be fickle with making mud textures too watery OR too dry. It can't get the perfect mud texture, but it's PRETTY GOOD when given the right inputs. Huge points for its multi-shot and physical interaction.

- Grok Imagine doesn't get what I want with my input. I will defer to SD77's experience with working with that specific mud texture, but it loses out for me due to its lower interaction and dynamic movement.

- Wtf is WAN 2.7 doing. It's an improvement over its previous models, but to avoid its notoriously bad visual artifacts, it makes everything look plastic. And it takes 10x as long to generate a video of the same length.

Based on what I'm seeing so far from our testing:
- Seedance 2.0 has perhaps the best potential if using text-to-video -- but it is also the most expensive
- Grok Imagine is probably the best value for simple outputs
- Kling 3.0 is closest to TV-quality style scenes

Which one is best depends on specifically what your use-case is and exactly what you're looking for in a QS scene. Seedance is simply too expensive even if it could work with good QS textures, and hasn't displaced my preference for Kling, which has always been strong with its physical interactions, and is now far easier to weave into long-form videos. Grok is an economical model to animate images.

[Viridian]

I wanted to see if this was a fluke, so I ran this as a text prompt: "A buxom jungle girl wearing a tattered loincloth is stuck in thick, heavy mud. Her struggle to escape the mud causes her to sink further into it." All I can say is that this is not the same model I tried out a week ago.

Oh, those are pretty impressive for text-to-image. It does emulate a muddy scene pretty well. That said, my use-case is constructing longer scenes by stitching these outputs together. It seems to be cohesive enough when it makes its own "rules" with a text-to-image prompt, but I suspect that it might do poorer with image-to-image.

[Sekani]

I wanted to see if this was a fluke, so I ran this as a text prompt: "A buxom jungle girl wearing a tattered loincloth is stuck in thick, heavy mud. Her struggle to escape the mud causes her to sink further into it." All I can say is that this is not the same model I tried out a week ago.

Oh, those are pretty impressive for text-to-image. It does emulate a muddy scene pretty well. That said, my use-case is constructing longer scenes by stitching these outputs together. It seems to be cohesive enough when it makes its own "rules" with a text-to-image prompt, but I suspect that it might do poorer with image-to-image.

I'll probably do some more testing later, but at the moment I'm just not sure what to do with a model that produces better results from text prompts than images. If there's a reliable way to get some kind of consistent output then Seedance might be worth it for me even at the higher cost (should note that it already costs less than it did a week ago). I'm willing to pay a little extra if I think that I'll have fewer unusable outputs.

As for Wan 2.7, I can't find a use case for it at all. Unfortunately these new releases will make 2.5 even more rare, and I expect that it'll be phased out before a suitable replacement comes along. Thankfully I'm amazed at what Kling and apparently Seedance will let you get away with if you don't specifically ask for it.