07 / Author's original manuscript / Practical notes · Observation records
From pouring coffee to large camera moves, preserving the author’s test notes and critical observations.
After nearly two weeks of experimenting and spending several thousand compute credits, I have finally put this major update through most of its tests.
To be honest, friends who do video post-production or commercial storyboarding must have a deep understanding: now watching the official demos of major AI videos, they have basically reached the level of "desensitization".
The official promotional videos always have perfect lighting and smooth camera movements, but as long as you hand-write prompts, you often still can't escape the same old problems - the screen will start to melt after two seconds, the character will directly "genetically mutate" as soon as you push and pull the camera, and pouring an extra glass of water will turn the cup into a transparent gel.
I focused on Kling 4.0 mainly because a few friends who make commercials privately told me that it had changed substantially in camera control and physical continuity. To avoid staying inside an information bubble, I took several real project shots that Runway and Sora had put me through the wringer with and ran a round of hellish stress tests.
Here is what it actually felt like to use.
First compare this real 4K coffee-pouring shot, watching how the liquid, ice cubes and glass actually make contact.
1. Before Image Quality, Does It Understand Basic Physics?
Today’s video models can produce single frames worthy of wallpaper, but once things start moving, you can see whether they are ‘artificial stupidity.’ What gave me the biggest headaches in earlier versions and similar tools was the collapse of physical laws when several forces interacted.
This time I tested a particularly tricky everyday shot: ‘Hot espresso slowly pours into a glass filled with irregularly shaped ice cubes, while the camera dollies in extremely slowly.’
Anyone who has tried video generation knows this combines fluid collisions, refraction through ice, translucent materials and dynamic splashes—a classic failure hotspot. Many earlier models made the coffee look like a floating lump of brown jelly; when it met an ice cube, it could even pass straight through it, with no surface tension to speak of.
Kling 4.0’s processing this time is indeed a bit surprising to me:
When the liquid strikes the edge of an ice cube, you can see tiny streams splitting and wetting the surface; its downward gravitational acceleration remains continuous.
The volume conservation of liquid accumulated at the bottom of the cup is calculated very accurately, and there is no weird bug where the liquid flows more and more, and finally overflows the mouth of the cup but has no source;
Most impressive are the refracted highlights in the backlighting. As the liquid level changes, the refraction angle shifts adaptively in real time, without a stuttering flicker every few frames.
In multilayer cloth simulations such as ‘sprinting through strong winds in a silk cloak,’ it finally escaped the clipping problem where clothing corners repeatedly sank into the thigh geometry. Even at a moderately high motion setting, the extending fabric folds still followed the wind direction. This suggests that underneath, it is doing more than simply guessing pixels from reference patterns: it is trying to build a relatively self-consistent system of physical logic across space and time.
Coffee Shots: Keeping Records That Can Actually Be Compared
The coffee pouring below is a real shooting reference, used to observe the relationship between liquid contact with ice cubes, glass edge and backlight, and is not a sample of the pressure test described in the original manuscript.
Fixed input
Keep the same opening frame, the same composition, and a clear pour direction.
If the entry provides a random seed, record the actual available fields.
Observe Each Dimension
Record the liquid contact, cup stability, liquid level rise and reflection flicker respectively to avoid generalizing all dimensions with "very real".
Keep the Failed Attempts
Keep failed versions and the differences between their prompts. The experience figures in this article come from the author’s own records; when reproducing the tests, use statistics from your own samples.
2. Decoupling Camera Motion: A Camera Turn Finally Stops Replacing the Person
In actual business storyboarding (Pre-vis), the word we are most afraid of is "semantic drift".
You want to take a cool 360-degree surround shot or a quick whip pan, but as soon as the camera movement command is issued, the AI often regenerates together with the subject, turns behind and back again, and the clothes and even the face of the protagonist completely change.
Kling 4.0 significantly strengthens the decoupling control of camera trajectory and subject characteristics.
For example, I tried an action scene storyboard: let the character punch and sprint in the rainy night alleyway, while the camera swooped in a wide arc.
In the past, high-speed motion blur was often the "fig leaf" of AI, but it was also the easiest to break down.
According to actual measurements, even if the picture is subject to strong radial blur when zooming rapidly, when the lens is refocused on the character, the location of the clothing material and mud and water stains are still exactly the same.
This means that its spatial three-dimensional perspective is maintained quite firmly.
For teams that want to use it to create storyboards and dynamic storyboards, at least they don’t have to use lifeless static shots for the entire film in order to preserve the characters’ looks.
This public action demonstration is used to supplement dynamic observation and is not the output of the author's coffee test.
3. How to Get Started and the Hidden Traps: Don’t Waste Your Compute Credits
Many friends ask how I usually run these test cases. If the goal is to meet a project deadline or explore the limits of prompts, a local setup with ComfyUI nodes certainly offers a high ceiling, but the cost of tuning parameters and trial and error can be daunting. For a quick check of the results, I personally tend to use the Kling 4.0 creation portal. This mainly saves the fiddly environment-configuration steps so I can focus on seed variation and the motion-intensity slider, with relatively direct feedback on the output.
Here is a prompt structure I developed over the past few days that produces usable results fairly consistently. Feel free to apply it directly:
[Explicit shot size and camera movement] + [Core subject action] + [Interaction between lighting properties and environmental materials] + [Details of force response] + [--motion 5-7]
A reminder: do not assume that pushing Motion intensity to its maximum is best.
Many people start at 8 to 10 to chase dramatic motion, and the image becomes very prone to breaking down. My experience is to keep motion intensity between 5 and 7, rein in the character’s movement a little, and leave the dynamism to camera movements in the prompt, such as Fast Dolly In or Low-angle tracking. Let camera displacement create tension in the image, and the result can instantly rise from cheap effects to a cinematic film feel.
This real portrait-format coffee-making clip shows another camera position and shot size, useful for designing tests at different levels of difficulty.
4. A Necessary Reality Check: It Is Not Perfect, and These Situations Still Fail
Since it is a hard-core test, there is no need to boast about it. Its shortcomings at this stage are still very glaring. Please try to avoid them when using it:
Stiff micro-expressions in macro close-up: As long as the camera reaches the level of facial pores, once the characters laugh, cry or recite lines violently, although the skin texture and eyeball reflections are extremely delicate, the muscles around the corners of the mouth and eyes still have a vague "silicone dummy feel".
Complex cross-occlusion of multiple people is still prone to adhesion: Once more than three subjects are crammed into the scene and they run and interlock with each other, the edge outlines will sometimes still have slight pixel adhesion, making it look like two people are wearing a corner of the same piece of clothing.
High-precision interaction (especially hands): If the character just runs, jumps, and waves, it's totally fine; but if you ask the character to pick up a bunch of extremely thin keys on the table, or count money quickly, there is a high probability that you will still see instantaneous distortion of the shape of the fingers.
5. Final conclusion
Overall, Kling 4.0 is no longer a toy limited to landscape establishing-shot wallpapers or animated PowerPoint backgrounds. Its advances in complex physical responses and controlled camera paths are enough to support animated advertising prototypes, visual-effects previs, and even finished online content for some low-budget projects.
If the motion stability and camera-related clipping in your current video project are wearing you out, try the Kling 4.0 creation portal yourself with a few difficult storyboard shots. It may not eliminate every bit of post-production work needed to repair clipping, but in video generation it has certainly delivered results that practitioners can place among the industry’s current leading tier.
It is more helpful to go to the test with a specific problem than to read ten beautiful examples. video creation reference portal can serve as a starting point for the next round of preparation.
Top comments (0)