Quicky Sanders: Bad Date (Kling 3.0 RELEASE & Review)

[Viridian]

Quicky Sanders arranges a Valentines date with a suspected criminal. She is lured into a deadly quicksand trap on Lover's Beach, and not by her date...

DeviantArt link: https://www.deviantart.com/viridianqs/a ... 1296010101

jaosiwdj.jpg

By coincidence, Kling released its 3.0 model just as I stated the first scene. I have used Bad Date as a test video for Kling 3.0. My thoughts are below.

THE GOOD:

- Multi-shot + Custom multi-shot.

Kling isn't the first to include multi-shot functionality, but it is perhaps the best now. It generally has more consistency with using the scene to generate appropriate cuts on its own. Most of the dialogue in this scene is auto-generated multi-shot, my favourite being Charlie's interaction at the end. This would normally have required many key frames to create a narrative, now all done automatically and cohesively.

Custom multi-shot allows each individual segment to have its own prompt. The starting scene with Sanders walking and close-ups of her chest and heels are examples of what can be specified.

- Up to 15 second generations + 1s adjustments

Previously, only 5s and 10s gens were possible with 2.6. Now, you can select anywhere between 3s and 15s, making for very adjustable generation lengths. This makes it work very well with the multi-shot function.

- Voice lines and expression

Kling continues to improve its voice delivery and expands language support (not used here). 3.0 clearly has an edge in its expression, and is reaching cinematic quality with its inbuilt multi-shot.

- Native Audio

Improvement over 2.6, generally better at matching visuals and audio. Notably, it has largely removed the annoying 2.6 problem of playing unwanted music for camera movements (e.g. not doing the suspense echo when zooming in).

- Smoother animation and interaction

2.6 struggled with "slo mo" and freezes; 3.0 has some lag time with dynamic movement but generally it has attained a realistic amount of movement, making this the most desirable model.

THE BAD:

- Tics and seizures

Outputs are inconsistent. Sometimes it will over-react to prompts showing suspense and anxiety, with characters twitching their bodies and faces as if they have Tourette's.

- Bad voice lines

3.0 sometimes struggles with delivering specific words for no reason. In this video, many instances of "quicksand" had to be removed because they continually said "quickSKWAND". Not a typo. It just forgets how to pronounce basic words.

- Last-frame voice lines

Often inserts voice lines at the very end of a generation, no matter how long the segment is. Results in last sound being in the very last frame, making it difficult to edit.

- Audio glitches

Still has blips of unwanted sound artefacts at the beginning of a generation, but has massive problems with inserting random noises during dead air. I consistently had Kling insert a laugh track at the beginning of a dialogue, and in one case included a trumpet and beatboxing out of nowhere.


THE UGLY

- Bad multi-shot output

When it works, especially for general scenes, it's amazing. But setting of up specific interactions is horrendously difficult. Kling also often generates entirely different scenes in multi-shot cuts rather than correctly using the elements in the key frame.

- Cost

3.0 costs a bit more than 2.6, but can save costs if generating in 720p. However, a 15s generation can be very costly given that if any of the above glitches happen in any frame, the entire thing is almost unusable.

MUD/QUICKSAND SPECIFIC

Can Kling 3.0 handle mud? In short, no.

It's better than 2.6, especially in the dialogue scenes. It generally does recognise the mud texture and keeps appropriate physical interaction. It can even handle controlled sinking and even submersions.

It is inferior to WAN-2.5 in its actual dynamic movement in mud, especially given its issues with seizures and tics, and extended mud interaction is more likely to result in water. If the source image has clear texture, 3.0 will work with it, but I don't quite trust it to struggle or sink.

There are opportunities to add immersive environmental interaction with multi-shot.

Overall, it's better than 2.6 and can function as an exclusive model for mud and QS scenes, but it's not trained with those interactions specifically and isn't as good as WAN-2.5.

OVERALL VERDICT

Kling 3.0 is a general improvement over Kling 2.6. It has more functionality, especially in its multi-shot, but it often has massive AI brain farts. The peaks are higher, but the failures are also lower.

I am definitely substituting Kling 3.0 for my cinematic scenes over Kling 2.6, but I will supplement it with other models that have more specific interactions.