꧁༺ 𝓝𝓪𝓻𝓽𝓶𝓸𝓿𝓮.𝓝𝓯𝓽 ༻꧂

1/2 🐉 WHAT IS HAPPENING WITH KLING AI? Why has China’s video-generation monster gone so quiet? Kling may look unusually silent from the outside. But inside Kuaishou, the company appears to be working on something bigger than just the next text-to-video model. On July 15, Kuaishou engineers published a paper introducing the Large Processing Model — LPM. This is not Kling 4.0. And it is not a new video-generation model. LPM is a generative system designed to take an existing video and reconstruct its quality. In simple terms, it is an extremely powerful AI video restoration model. 🛠️ What LPM can do: 🔹 Remove noise and compression artifacts 🔹 Recover fine visual details 🔹 Improve faces, hair, clothing, foliage and architecture 🔹 Preserve text and subtitles 🔹 Reduce flickering between frames 🔹 Maintain consistency across long videos A traditional upscaler mainly increases resolution and sharpness. LPM uses a diffusion-based generative model to reconstruct visual information that may have been lost during compression, recording or transmission. But this creates an obvious danger. A generative model can start inventing things. It might change letters on a sign, redraw a face, replace a pattern on clothing or generate slightly different details in neighboring frames. So Kuaishou specifically trained the system to leave already-correct regions untouched. ⚙️ HOW IT WORKS The project began with LPM-Image, a restoration model for individual images. To train it, Kuaishou created the Kwai UltraVision Dataset, containing billions of visual examples, including complex textures such as: skin, hair, fur, foliage, fabric and architecture. The system was then adapted for video. And video introduces a much harder problem: temporal consistency. It is not enough to restore one frame beautifully. A mole, a face, a fabric pattern, lighting and background details must remain identical throughout the entire video. Long videos are especially difficult. They normally have to be divided into smaller segments. But errors introduced in one segment can gradually propagate into the next. That is when colors, faces and textures begin to drift. To solve this, Kuaishou introduced Temporal-Pyramid Inference. The system first generates sparse anchor frames across the entire video. Then it adds intermediate anchors. Only after that does it restore the remaining frames. It works almost like a keyframe structure: first, the model establishes the overall visual framework; then it fills in the gaps. This reduces accumulated errors and allows the system to process arbitrarily long videos with much less visual drift. 💥 AND HERE IS THE MOST IMPORTANT PART LPM is not just a laboratory experiment. Kuaishou says videos processed by the system now account for approximately 45% of total viewing time on the platform. The company also reports that LPM can reduce bitrate by around 20% while maintaining comparable perceived visual quality. That means a video can be compressed more aggressively, delivered faster and more cheaply — and then restored by a generative model on the other side. This is not only about prettier video. It is also about infrastructure, bandwidth and production cost. 🎬 WHAT DOES THIS HAVE TO DO WITH KLING? The paper explicitly states that LPM’s low serving cost makes it suitable for integration into products such as Kling. Kuaishou does not disclose exactly where or how it is already being used, so everything beyond that point is an inference. But LPM could potentially handle: 🔸 Final enhancement of generated clips 🔸 Detail reconstruction 🔸 Flicker reduction 🔸 Higher-quality 4K upscaling 🔸 Stability across longer videos 🔸 Better preservation of faces and text 🔸 Lower generation and delivery costs In other words, Kling could generate video at a resolution that is more computationally efficient — and then send the result through a separate restoration model for the final pass.

🚀 Google I/O 2026: What I’m expecting on May 19–20 In three weeks, Google will hold its biggest conference of the year. The keynote starts on May 19 at 20:00 Moscow time. Based on the rumors and insider signals I’ve been collecting over the past few weeks, this could be one of the most packed Google I/O events in the company’s history.