Kling 2.6 Animation Update

[Viridian]

In the past two days, Kling has seen some HUGE updates.

The first is Kling Omni 1 (O1). This is the biggest interface update yet. In short, O1 is meant to replace the manual prompting by creating a chat-bot style interface, so you can describe your image or video in normal prose and Kling will make the adjustments. This has been very powerful - I've been using this to extend my source images. For example, I might general a cartoon image and get it to make it photorealistic, or extend a background, or change an object or costume. Notably, this can work for both images and visuals, so you can smoothly integrate any prompt.

For example, in this recent commission, the brief included a temple and the character.
https://www.deviantart.com/viridianqs/art/Jinxie-s-Temple-Commission-1271030422

I generated the temple flyover shot first using Kling O1, then added that as the source image along with the character as a second source image, and it made a new clip that made the character walk in the scene that it had created. I'm very impressed with the visual cohesion. This is been a massive upgrade to my workflow, as this automates the manual photomanipulation I had to do previously, allowing me to generate the backgrounds using plain English for my characters, and it will animate it too.

O1's video model is on par with its 2.5 Turbo release, so it's cohesive, though kind of floaty and sluggish, but the interaction with the elements is top tier. Most AI models are trying to reach for that function (VEO has been pushing it a lot). Kling generally has better interaction, though not perfect. Still, Kling O1 brought me back to it after WAN-2.5's incredibly mud physics - but only because WAN is mediocre at animating everything other than mud.

Today, Kling 2.6 released. This one is also a huge update. It builds on top of the 2.5 Turbo model. Where 2.5 is a "curated" feel, 2.6 releases some of its creativity. Its movement is the best blend of 2.5 and 2.1 - the movement speed is better, the physics more flowy (and jiggly) without being 2.1 over-the-top, though at a base level it isn't as good at element interaction as O1.

The biggest change is its native audio. Native audio has been a core function that set Sora 2 and VEO 3 apart, and expected (as we see in WAN-2.5). Kling's audio has always been laughable - the equivalent of someone recording a speaker on a wooden table.

2.6 has finally upped the game to make the audio on par with current gen AI. It's not great - VEO 3.1 and Sora 2 probably do better in that regard, but it's serviceable, so you benefit from Kling's dynamic interaction whereas VEO is janky.

It is capable of voice lines too, which is a huge improvement. Unfortunately, it's... not very good. It flubs its lines similar to how VEO 3 started, so there's a lot to patch up, but at least it has voice lines that are serviceable. The big gain for me is that I don't need to use VEO's awful non-functional platform and uncontrollable animation for voices. I immediately cancelled my VEO 3 subscription -- I'd rather put up with Kling's mid-level voice gen and better visual adherence than have VEO continually take over the scene with unprompted actions - that's assuming the VEO even works half the time.

Unfortunately, 2.6 doesn't seen an improvement in QS physics. It might be better in some ways, and it can pick up up the mud interaction, but it hasn't seen much improvement in recognising how the character is meant to interact with the mud. WAN-2.5 is undeniably the winner here with its QS physics, clearly being trained on QS scenes where Kling can do it, but isn't trained for it, and VEO just can't do it.

I've attached several samples of Kling 2.6 to show its physics and voice lines.

My opinion of current releases between the three main models I use (VEO, Kling and WAN):

Quicksand: WAN > Kling > VEO
General movement: Kling > VEO > WAN
Voice lines: Sora > VEO > Kling > WAN
General sound: VEO > Kling > WAN
NSFW Filter: WAN > Kling > VEO
User interface: Kling > WAN > VEO
Visual cohesion: Kling > VEO > WAN
Prompt adherence: Kling > WAN > VEO
Creativity: WAN > Kling > VEO

Tilt factors:
VEO has the best control over dialogue, but is falling behind in animation and adherence.
WAN has superior mud interaction, and uniquely can be used to generate an entire scene including the initial off-screen entrance into mud (the other two need a starting source image) -- but WAN warps BADLY even in the same output.
Kling is generally the most balanced in its strengths and weaknesses, but the Omni 1 interface gives it a huge amount of flexibility that can actually replace image generators.

[Sandy Place]

Interesting. Nice work, btw.

_3 is my favorite, even without her uttering the magic word.

[H78capture]

Great videos, thanks for posting!

[Viridian]

Test run of Kling O1 / 2.6. Apart from the character references I used and some background tracks and waves, EVERYTHING in this scene is generating from Kling. This is the beginning of my magnum opus - I'm aiming to create a 10+ minute quicksand film.

[Sandy Place]

If this, um, body of work is any indication, I predict you will be quite busy before long.