Quicksand Fans

...and at least for the subject of thick sloppy mud scenes, it is by far and away the best so far. Your prompt can include the words "mud" and "sinking" and it will create a sinking video, without it turning into a watery mess. It's the first model I've found that can do a character sinking deeper step by step as they walk (Kling 2.1 tended to make the character sit down, or suddenly fall into a hole of muddy water).

It's slightly more expensive (50cr) than 2.1 standard (which is 40cr), but cheaper than 2.1 pro. It's only available in "pro" mode. Given it usually takes several takes to get the video you want, in the end it works out a fair bit cheaper because you don't have to generate so many videos refining your prompts or starting image.

Unfortunately, 2.5 Turbo gets an "OK" grade for me, in that it's a significant downgrade compared to 2.1 (Pro). But I'm also not the typical user. I'm one of the few, if not only, AI creator making extended cinematic-style scenes which involves several different tools to make the scenes work.

My general feeling with 2.5 Turbo, as the name suggests, is that it's a curated, streamlined version. It's not necessarily "better". It does select from its training set the most consistent outputs, so your result will be more cohesive. That means it will follow the prompt more closely, maintain the visuals more consistently (less warping) and have fewer AI "tics".

However, this comes at the cost of dynamism. The animations are OK, but they have floaty, sluggish feel that is typical of AI animation (compared with VEO and Runway). The character animations feel stiff and wooden, as if the model is trying to restrain the character so it doesn't fail to adhere to the prompt, but in doing so it is less creative in interpreting the prompt.

I've attached examples of my outputs, comparing 2.1 Pro with 2.5 Turbo (which is Pro by default).

The "Flash Gordon" Run

2.1 has crisper, more action and real-time movement, and the physics reflect that through the hair and clothes. It _is_ janky - because there's so much rapid movement, the AI can at times have jitters in strides, arms and legs, etc. It will occasionally have weird interactions with the environment it creates (e.g. if you say that the character jumps over a log, the character will often just jump in place and a log appears elsewhere). 2.1 has a major downfall in that if you don't provide enough prompting, it runs away with the visual input - one of the biggest problems (still) is the character moving backwards or turning around and running the other way. You _have_ to prompt for camera movement to nudge it in the right direction, and that's probably what most creators are failing to do.

2.5 is a much cleaner run. It doesn't have issues with environmental interaction and the animation looks less jittery. There's less wrong with it. It's just that it looks... boring. I've made dozens of Kling videos and my work with 2.1 is still pushing what it can do on cinematic level, so 2.5 looks like a downgrade in its movements. It feels barely better than 1.6.

The verdict here is that 2.5 is better at giving the output you want, and 2.1 is wilder.

The Tangled Walk

This scene has the Rapunzel character walking nervously through a dark forest.

2.1's version displays the typical advantage (and fault): it is much more creative in interpreting the visual and textual prompt. When it reads "nervous", it moves the camera and makes the character express it through facial and body movements. This makes for a much more evolved output. However, this two-edged sword: often you don't actually want these extra actions (sidenote: characters use their arms way too much when talking expressively, even when they're buried in quicksand), and this also introduces a noticeable amount of warping as characters and costumes gradually change as the clip progresses. I often have to do a "reset" shot in my videos so I can regenerate the still image correctly.

2.5 keeps the character's appearance more consistently, but it doesn't do much else. The walk is fairly standard. I feel that you have to do a LOT more prompting, which isn't a bad thing if you're really into prompt engineering, but it's more arduous to prompt for exact movements for each body part and environmental item.

Now let's do some spicier comparisons.

Jiggle Physics

Let's be honest, this is probably why Kling is appealing to a lot of people. Based off the numerous complaints I've been seeing on DA and on the Kling discord, this is what people are most upset over, mostly because the censor is starting to run away by blocking things on 2.1 that were OK a week ago (I've been been able to do nip slips and even toplessness, but not now). 2.5 is even stricter in how it understands the filter, so fewer things will pass through. I've had tight clothing or hard nipples blocked entirely.

The comparison here isn't even close. 2.1's "jiggle" physics are superior, but aside from that distraction, the body movements feel more dynamic. Yes, it does slowly introduce some warp and it can feel unnecessary and unwanted, but this is more of a visual prompt thing. 2.1 is amazing at interpreting what to render dynamically. A more covered woman would not have much "jiggle", but if it sees that the character is not wearing a supportive bra (e.g. deep cleavage), it goes anime-style on it. But this is often more realistic.

2.5 tones it down by... a lot. And this does result in a scene that is more consistent in its style, but it feels stiff and unnatural. It's actually not bad, but when you're used to 2.1's movement, 2.5 looks weird. It's not "off", but for all the examples I've seen of people praising its movements, 2.5 looks worse compared to 2.1.

However, the biggest advantage is that 2.5 has superior Text to Image, whereas 2.1 doesn't support it (2.1 Master does, but it's the top price tier).

Island Shot

Since 2.1 can't render Text to Image, I'm comparing 2.5 with its direct rival, VEO 3 (Fast). While the sign-up cost for Google AI is fairly pricey, the outputs are comparatively cheap. I'm on the Google Ultra plan (discount for first three months), and not only do you get a TON of credits to work with (with choice to top of), Ultra users also get FREE VEO 3 Fast outputs. But's VERY expensive to subscribe, so it doesn't balance out if you're not a power user.

Anyway, it seems that 2.5 is trying to vie with VEO in this department. The environmental outputs are about the same. However (not shown in these samples), Kling is already substantially better in its movement, while VEO 3 is stuck with video-game stiffness of animations. If you're creating a scene from text only, Kling 2.5 is significantly better.

However, Kling's audio is notoriously bad. It's improved recently with the option to add prompts for sound, but the quality is very poor, often cuts out sometimes generates sounds like someone bumped into a table, like everything was recorded in a microphone. VEO 3's sound output is currently matched, though there's far less control and VEO tends to run away with adding things like background music and it REALLY struggles with line reading. Recently I've had about a 30% failure rate where characters just speak gibberish lines instead of following my prompt. Additionally it has trouble NOT being creative - it ties to actually create the scene based on the line you give it rather than making the character just say the line.

Dumb example I experience: I had a character say "Now all three of us are stuck", and VEO actually made two more characters appear in the video. Obviously unusable.

Now the quicksand parts. I use Image to Video for all of these.

Struggling

This actually is a case of "less is more". 2.1 has far more dynamic movement, again. It does more to interact with the scene, sometimes too much. The texture and physics are amazing (especially compared to 2.1), but you most know how to prompt it. I generally like these interactions better, but they might not be what I need, so 2.1 is a coin flip. The outputs are typically very good, but not what I need. Consequently I might chop up footage and use all of them but in a different order, so it's salvageable.

2.5, in this case, actually has an edge because it doesn't do as much with the scene. The character moves less, and in doing so preserves the visuals better. I can see myself using 2.5 for specific scenes like this. However, I did note that 2.5 tends to be more literal in its interpretation of sand and mud. I find it generated sand out of nowhere often.

Sinking

Hold up, this completely changed my mind about 2.5. 2.5 is outright better.

Here's the nuance: 2.1 is actually very good already, but it strongly depends on camera prompting. In this example, I would have to specify the camera pans down and zooms in, which would nudge the AI to make the characters sink down. She does in this case, but it tries to do a bit more with the scene - the scene included a "speaking" instruction, and 2.5 tends to ignore that.

2.5 doesn't f--k around. You tell it to sink, it sinks.

My verdict

2.1 Pro > 2.5 Turbo Pro > 2.1 Standard

2.5 Pro: More adherent, more consistent, less dynamic, less creative
2.1 Pro: More dynamic physics, less consistency (higher error rate, more jittery)

I don't compare 2.1 Master. 2.1 Master is good, but the output isn't 5x better than 2.1 Pro to make it worth the cost.

The caveat is that I do far more editing to make 2.1 scenes work. It's like riding a wild horse compared to a local pony ride. I can see where 2.5 might be useful in specific scenes where you don't want a lot of dynamic movement and where 2.1 goes a bit too far.

Let's not overlook that 2.5 Turbo is a _cheaper_ model compared to 2.1 (Professional). If it's objectively better than 2.1, i don't think it would have been priced lower. I believe this reflects the trend of users wanting options for cheaper outputs rather than better AI animation, and to make it compete directly with VEO.

The censor with these things is undoubtedly a problem. Things like "running shorts" and nothing even in the slightest bit risqué will trigger it (although sometimes rearranging the prompt will stop it). "Thighs" will often trigger it in completely non-sexual contexts (e.g. "His powerful thighs work hard as he climbs the cliff face"). Good luck getting it to ever accept "lycra shorts".

An update on my experiences with 2.5 Turbo. I've been trying to weave more of it into my work to really see if it's better - and partially because it's cheaper.

I may have been harsh on my initial judgement, but my sentiments remain. 2.5 Turbo is less expressive, but it is substantially more stable. It does better are rendering more conservative scenes, but at the same time occasionally has more significant AI "brain" farts.

In this regard, I'd say 2.5 Turbo is easily superior in maintaining the cohesion of the mud. As mjw has experienced, it's far less likely to turn it into water, whereas 2.1 does struggle substantially if a scene does contain something resembling water, and even shuffling off-screen, it can introduce puddles.

The downside of the superior texture is that 2.5 has a harder time expressive dynamic movement and emotion. Facial reactions are robotic and plastic compared to 2.1 And, as shown above, the comparison between dynamic movement is in favour of 2.1, though 2.5 has more consistent outputs.

Currently I'm flicking to 2.5 when I feel like I need a guaranteed output that I can use rather than gambling with 2.1. I use 2.1 for "safe" gambles such as walking scenes, but 2.5 for "riskier" renders such as introducing new elements or working with mud textures that require specific movement.

2.5's biggest edge is that it response far better to prompts that specify depth. For example, you can tell it to sink a character to waist our chest-depth and it is far more likely to actually do it, whereas 2.1 will plunge the character through to the next dimension. I've only made 2.1 look like it adheres by using edits

That feeling when I can't use Kling at all since I don't have subscription.

One thing I've noticed with 2.5 is that if you're not careful, it can tend to cause faces/people to morph, so the person at the end of the video isn't really quite the same as the one at the start, something I've not seen happen with 2.1.

I've definitely noticed this happening with both. Considering that I make 5+ minute video where I use the last frames, I see a significant amount of warping between each clip. I find 2.5 tends to be a little better at preserving the original composition of the source image, which is why it actually does quite well with mud textures. 2.1 strongly processes this as water, whereas if you say it's mud, 2.5 will use a mud texture.

Minor update: it seems that Kling is addressing the censorship issues. I've managed to run previously OK / then blocked / now OK items. Not 100%, but thee are definitely changes in the filter happening live. I feel that I'm currently able to make 80% of the style of visuals that I previously made, though there are still dumb censor triggers, notably in how it reads written prompts.

For those who aren't using Kling and want a snapshot of the comical aggression of the censor, it over-reads fairly innocuous words like "her hands" or "her legs", as it was trained on sexual content, but it might be as straightforward as "she holds her hand up" (BLOCKED!). Changing the word order, removing the line and so on makes the video render fine.