NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Kling 2.0 Adds Image-and-Voice Editing, Shifting AI Video From Generation to Conversation

Kuaishou's Kling 2.0 introduces multimodal visual language so creators can edit AI video with images, voice, and motion instead of prompts alone.

2026-11-13 · 959 words · NeuroAI
Kling 2.0 Adds Image-and-Voice Editing, Shifting AI Video From Generation to Conversation

Most AI video tools still work like a vending machine: type a sentence, get a clip, and hope it matches the idea in your head. Kuaishou (快手) decided that was the wrong metaphor. At a Beijing launch event in April 2025, the company pushed its Kling AI (可灵) system into a 2.0 generation — and the headline feature was not "better clips" but "editable clips."

Kling 2.0 is a video-generation large model (大模型) built for a different job: letting a creator steer the output after it exists, using reference images, voice cues, and motion paths rather than rewriting a text prompt from scratch.

What actually changed in 2.0

The upgrade arrived as two models at once. Kling AI 2.0 covers video generation, while the companion Kolors 2.0 (可图) handles image generation. The shared idea is a concept Kuaishou calls Multi-modal Visual Language (MVL).

MVL splits human intent into two parts:

  • TXT — a "semantic skeleton" of plain text describing the scene.
  • MMW — "Multi-modal-document as a Word," a way to drop an image, a video clip, or other signals directly into the prompt as if it were a token.

In practice, that means a creator can show the model a product photo and say "keep this subject, replace the background with a studio," or hand it a reference clip and ask it to add, remove, or replace elements inside the generated video. Kuaishou describes this as multimodal (多模态) editing, and it is the clearest break from the single-prompt workflow most tools still use.

Kolors 2.0 arrived alongside it with its own controls: partial repaint, outpainting, and style transfer across more than 60 stylized effects.

The scale behind the launch

Kling was not a quiet research release. By the April 2025 event, Kuaishou said the system had passed 22 million global users, roughly ten months after its first public launch in June 2024, and had gone through more than 20 model iterations.

Before 2.0 shipped, an earlier version had already drawn outside attention. On Artificial Analysis' global video leaderboard in March 2025, Kling 1.6 Pro (high-quality mode) reached an Arena ELO of 1,000 in the image-to-video track, ahead of Google Veo 2 and Pika. That score belongs to the prior version, not 2.0, but it set the competitive frame Kuaishou was arguing from.

On its own internal tests, the company made a stronger claim. In Kuaishou's GSB (Good-Same-Bad) evaluations, it said Kling AI 2.0 posted a win-loss ratio of 182% against Google Veo 2 and 178% against Runway Gen-4 on semantic responsiveness, visual quality, and dynamic motion. Those numbers are Kuaishou's, not an independent lab's, and should be read as the company's positioning rather than settled fact.

Where Kling sits in the video race

Kling is not alone. By 2025, video generation had become one of the most contested fronts in large-model (大模型) development, with closed systems from US labs and a fast-growing set of Chinese models competing on quality, motion realism, and editing control. Kuaishou's advantage is unusual: it is a consumer platform first, so Kling is trained and tested against the daily needs of short-video creators rather than only benchmark prompts. That gives it a built-in distribution channel — the same app where creators publish is where they can generate. The risk runs the other way: a product tuned for social clips may not translate cleanly to film, advertising, or education without further work. Kling 2.0's editing layer is the company's answer to that gap, shifting the value from "make a video" to "shape a video you already have."

Why Kuaishou is betting on editing

The strategy points at a real pain point. As Kuaishou senior vice president Gai Kun put it at the launch, text alone is an incomplete way to describe an image, and the industry still falls short on the stability and precision creators actually need. Kling 2.0's answer is to treat the first generated clip as a starting point, not a finished product.

That matters for the business. Video generation only becomes useful at scale if teams can iterate quickly, not if every change means writing a new paragraph of prompt. Kuaishou has been leaning into exactly that: its 2025 full-year results show Kling AI brought in about $47 million (RMB 340 million) of revenue in the fourth quarter alone, and that December 2025 monthly revenue crossed $20 million, an annualized run-rate near $240 million. The company also open-sourced a 671B-parameter multimodal large model called Keye (快意) in late 2025, signaling it is building both closed creative products and open research weights.

Honest limitations

Kling 2.0 is a commercial, closed product, so independent researchers cannot inspect its weights or reproduce its benchmark claims. The 182%/178% win-loss ratios are Kuaishou's internal GSB results, not a neutral evaluation, and GSB methodology is sensitive to how prompts and raters are chosen. Like every video model in 2026, Kling still struggles with long-form coherence, consistent characters across shots, and physically plausible motion over more than a few seconds. The multimodal editing tools are powerful for short social and ad clips, but they are not a replacement for timeline-based post-production on longer narratives.

What readers can do now

  • Try the editing workflow before judging the quality. Sign up for Kling AI and test one short clip with an image reference plus a text edit, not a from-scratch prompt — that is where 2.0 differs from 1.x.
  • Watch the benchmark, not the press release. Track Kling on Artificial Analysis' video arena over the next quarters; an independent ELO trend tells you more than a launch-day GSB ratio.
  • If you build with it, budget for API costs and test consistency. Run a small batch of same-subject clips to see how stable characters and objects stay before committing to a campaign.

Related coverage

Cite this article

NeuroAI (脑机网), "Kling 2.0 Adds Image-and-Voice Editing, Shifting AI Video From Generation to Conversation," 13 November 2026, https://neuroai.site/a/na-model-kling-2-video-gen

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories