Author | Gao Xiaoyang

Editor | Hao Xin

On October 15, Kling 4.0 was officially rolled out to all users. Kling has encapsulated the essence of this upgrade as “realism, control, and professionalism.”

In terms of features, Kling 4.0 supports generating videos of up to 30 seconds in a single pass, with up to 10 flexible keyframes. The prompt input limit has been raised to 8000 tokens, and it now supports the 21:9 aspect ratio and dual-channel stereo sound.

But that’s not all—upcoming features include 4K and 1080p 10-bit HDR output, as well as video extensions of up to 2 minutes.

However, more parameters don’t automatically translate to greater creative control. The real questions are specific: Can a chase scene maintain character consistency during continuous motion? Can a face retain its identity across different ages? Can dialects and spatial audio align properly with the visuals? And when editing a completed video, can only designated parts be modified?

To answer these questions, we designed four rounds of hands-on tests to see if Kling 4.0 truly delivers on its promise of “boundless imagination, instant inspiration-to-film.”

30-Second Snowy Night Chase: Long Takes Gain Continuity

Kling 4.0 raises the native generation limit to 30 seconds and emphasizes model stability in dynamic, complex camera movement scenes.

Using the snowy night setting of Millennium Harbin from Drunk Dream, we designed an old-town chase scene: a character bursts out the back door of a video hall, runs through icy alleys, and finally enters an overpass. The combination of chase, low light, and winter night allows us to observe continuous motion and dynamic image quality.

The challenge was fitting running, pursuit, and camera movement into a 30-second sequence. If the character’s face changed mid-scene or pursuers lost spatial relation to the protagonist, the long take would lose narrative coherence.

Kling 4.0 delivered on all fronts. The character’s hairstyle, clothing, and facial features remained consistent, with no mid-sequence swaps. Snow particles, ice reflections, and lighting layers were clear during high-speed tracking shots, with no heavy blurring in low light.

The camera smoothly transitioned from behind to in front as the character rounded a brick column, using body movement to mask perspective changes. Some actions were simplified—climbing over a bicycle became more of a crouch, and bursting through an iron gate became passing through an open door. Pursuers didn’t constantly stick to the protagonist, slightly reducing urgency but keeping the 30-second running main storyline intact without frequent jump cuts.

The most valuable change is the model’s prioritization of scene continuity. It didn’t follow every action in the prompt verbatim but preserved characters, settings, and main action lines—giving concrete meaning to “process coherence.”

10 Keyframes Span 60 Years: Control Comes from Clear Nodes

Kling 4.0 supports free keyframe placement, allowing up to 10 keyframes to specify character states, scene changes, and plot points.

Drawing on the decades-spanning performance theme from Orchid Fragrance As Before, we used the narrative of “a woman aging from 16 to 80” to test Kling 4.0’s input method.

The age-change requirement involved altering skin texture, hair color, posture, and expression without changing identity. More nodes let creators define each stage precisely, increasing constraints for the model. Providing only a young reference image simplified input but left age progression more to the model’s judgment.

Next, we tested which approach was more reliable.

We generated two 10-second videos: one with 10 keyframes (16 to 80 years old), the other with only a 16-year-old reference image, letting the model infer youth, middle age, and old age.

The 10-keyframe version fully adopted all inputs without disrupting the uploaded sequence. The character aged smoothly, with clothing and hairstyles matching each stage, and nuanced micro-expressions across ages. We pre-visualized each age with GPT, leaving the model to handle motion and emotion between nodes.

Since jumps occurred between frames, we let the model infer by uploading only the 16-year-old reference.

The single-reference version still had flaws—transitions remained rough. Early changes focused more on clothing than age, with a rapid shift from youth to old age. The model completed age progression but rushed intermediate stages, leaving facial and physical changes insufficiently gradual.

Each input method suits different control needs. Simple actions can be left to the model, but for core content, clear keyframes are more reliable. “Directorial sense” becomes concrete—creators decide key moments, while the model fills in transitions.

Dialects and Lip-Sync Hold Up, Spatial Sound Better for Post

Kling 4.0 upgrades dialogue and audio capabilities, claiming more accurate lip-sync, natural synchronization of sound, dialogue, and performance, plus dual-channel stereo support. Chinese accents cover Northeastern, Sichuanese, and Cantonese dialects.

We set the scene in a second-floor law office of a Northeastern bathhouse from the TV drama Prospering Every Day, generating the same contract consultation dialogue in Northeastern and Cantonese dialects. Dialects needed natural intonation and rhythm, with lip-sync matching longer lines to avoid unrealistic performances.

Both dialect versions maintained high fidelity without reverting to Mandarin, with word-for-word alignment between dialogue and lip-sync—no swallowed, mispronounced, or mixed-up words. For solo dialogues common in short dramas and ads, these results are directly usable.

We also designed a 5-second spatial sound task: placing a bathhouse worker’s shout on the left, a legal TV program on the right, and the female lawyer’s dialogue centered in the foreground. In testing, voices stayed synced with the visuals, but background shouts and TV audio weren’t clear, and left/right channel relationships were indistinct.

This suggests the model prioritizes on-screen subjects and dialogue, potentially overlooking complex ambient sounds. For sound-driven narratives, generating ambient sounds separately and mixing in post remains more stable. The most practical audio upgrade here is dialogue; spatial sound works better as a supplementary feature.

One-Pass Editing of Finished Shorts: Precision Edits Align with Real Workflows

Video precision editing brings Kling 4.0 another step closer from a generation tool to a post-production tool. Officially, it can add, modify, or delete subjects and backgrounds in the original video, as well as adjust style, weather, color, and texture.

We took an ~11-second “No Heartburn While Eating” AI short drama clip and asked Kling 4.0 to modify it, testing if its edits were as “hassle-free” as the steamed buns in the video. Finished clips like this better test editing than zero-to-generation, as unnecessary changes would disrupt the original.

We input the source video and three reference images, requesting removal of watermarks and English subtitles (keeping Chinese subtitles intact), addition of a middle-aged male customer reading a newspaper (replacing the original female customer), and a rainy-night cyberpunk style shift from daytime. Edits had to execute multiple new requirements while preserving original character actions and camera movements.

In testing, all four edit types were executed, and all three reference images were used. Platform watermarks and English subtitles were cleanly removed; Chinese subtitles remained unchanged in text, position, and font. Original camera movements were preserved, the replaced female character closely matched the reference, and her original position and speaking state weren’t altered by the swap.

What best demonstrated “precision” was the simultaneous removal and retention. The model didn’t treat all text as identical or damage adjacent frames while removing watermarks.

The results weren’t perfect—the newly added newspaper reader fused awkwardly with the original frame, and the background jumped during style conversion. But asking the model to simultaneously add/replace characters, remove elements, and convert overall style involves multiple constraints, making this a stress test. We can’t fault the model for not 100% meeting demands.

For official final cuts with Kling 4.0, we’d first complete character and text edits, then handle background and tone uniformly, keeping each round’s goals focused.

From Iterative Generation to Clear Control: Kling 4.0 Takes a Concrete Step Forward

Together, these four tasks show Kling 4.0’s progress isn’t just about longer videos or more input fields. The 30-second long take preserves characters and main action lines; 10 keyframes enable pre-planning for cross-age narratives; dialects and lip-sync jointly support characters; and video editing distinguishes what to change from what to keep.

These capabilities target the most time-consuming creative steps. Where multiple generations were once needed to find suitable characters, actions, and shots, now duration, keyframes, reference images, and edit commands can pre-define some outcomes. The model still makes choices, but a clearer division of labor emerges between creator and model.

Testing shows version 4.0 hasn’t yet reached “prompt-to-final-cut” simplicity. Complex actions may be compressed, autonomous inference is less stable than clear nodes, multi-layer spatial sound still needs post-supplementation, and composite edits work better as focused tasks.

For creators focused on short dramas, ads, and narrative clips, Kling 4.0 is now workflow-ready, especially for character consistency, key plot control, and modifying existing footage. For those mainly producing simple atmospheric shots, improvements are more about image stability and dialogue completion.

The real reduction here isn’t the few minutes saved per generation, but the shrinking gap between creative intent and final visuals. When creators can define which nodes must stay fixed and which materials must remain, iterative generation starts becoming purposeful production.