An AI system lets users change a video’s lighting as it plays

LiveLight responds to changing three-dimensional light positions, colors and intensities while generating video frame by frame.

Editorial illustration — not from the study.

LiveLight uses depth-aware maps to describe how a changing light source should illuminate a subject. A lightweight control module feeds those maps into a video-generation system, while geometry-based training helps preserve plausible shading when the system uses only a few processing steps.

To keep the video continuous, LiveLight generates frames in rolling chunks and carries intermediate information forward. This allows users to adjust a light’s position, intensity or color during generation rather than specifying the lighting for an entire clip in advance.

How the relighting works

The researchers report that LiveLight achieved state-of-the-art relighting quality at real-time speed on real-world and synthetic benchmarks. Compared with offline baselines, it performed better in temporal stability, lighting controllability and user preference. The system uses four denoising steps by default, with a reported speed of 15.78 frames per second; using eight steps raised quality marginally but reduced speed to 9.43 frames per second. The method can refresh its reference on each frame and is intended to support arbitrarily long streaming video.

Why live control matters

Most video-relighting systems described by the researchers require the lighting plan for a whole clip before generation. LiveLight instead allows lighting changes during streaming, so a user can adjust a virtual three-dimensional light source and see the resulting changes in the video as it is produced. That makes interactive lighting control available within a diffusion-based video system rather than limiting it to offline processing or portrait-focused tools.

Evidence and caveats

The evidence comes from experiments on real-world and synthetic benchmarks, including comparisons with offline methods and assessments of temporal stability, lighting control and user preference. The excerpts do not provide the benchmark sizes or detailed numerical results. LiveLight has difficulty with small or distant faces and can lose fine details or produce artifacts in text and small accessories. Its geometry-based guidance also depends on depth and surface-normal estimates, which can be inaccurate for complex hairstyles, occlusions and extreme poses. The researchers attribute some detail problems to the capacity of the underlying image decoder.

One video can animate many animal and human characters—about 20 times faster

Researchers developed an AI system that captures motion from a single video and transfers it to characters with different skeletons, including human and animal forms. On two benchmark collections, it reduced average joint-rotation error from about 17 degrees to about 10 degrees and ran about 20 times faster than mesh-based pipelines.

3 hours ago·3 views

AI & Computing

AI models shared human values—but not how strongly to rank them

Most of 75 large language models ranked six AI priorities in the same order as 376 human respondents. But some models overstated the gaps between those priorities, and closer agreement with people did not consistently come with greater size, recency or capability.

16 hours ago·7 views

AI & Computing

A client-side AI system blocked 1.2 million malicious email logins

A lightweight language model running on client devices detected brute-force and credential-stuffing attacks against email accounts. Across 2.5 million authentication sessions, it reported 98.4% attack-detection accuracy and blocked more than 1.2 million malicious login attempts during a six-month deployment.

23 hours ago·6 views