Video Diffusion News
The latest Video Diffusion coverage curated for game developers.
A new speech-animation pipeline, AnyTalk, can generate lip-synced 3D talking characters without character-specific animation data. It adapts a video diffusion model to a target mesh, then lifts the generated talking-head …
MV-Forcing tackles a long-standing gap in generative video: producing dynamic scenes that stay consistent across both time and multiple viewpoints. The system combines temporal and view-wise autoregression with a 4D …
ControlHair combines a physics simulator with video diffusion to make dynamic hair rendering more controllable. The pipeline turns simulated hair motion into per-frame control signals, then drives a diffusion model …
A paper shows a large video diffusion model can reconstruct Hitchcock’s Vertigo scene-for-scene from just 2.78% of the original frames. For game teams, the interesting part is how much cinematic …
A controlled study of action-conditioned world models finds that memory design matters more than replay quality suggests. With a shared video diffusion backbone and fixed action interface, the authors show …
A study of video diffusion models found they can linearly expose physical plausibility in their internal states, even though that signal is missing from the VAE input. On IntPhys and …
A new arXiv paper benchmarks whether models can infer 2D physical dynamics from video and then extrapolate them. For game teams, the interesting part is the tradeoff: code generation gives …
StreetNVS uses all three signals modern vehicles already collect—LiDAR, surround cameras, and ego-motion—to synthesize novel street-view video. The big shift is a reference-enhanced camera attention module plus a training curriculum …
A new paper says INT2 KV-cache quantization in chunked video diffusion can quietly bias attention, making cached keys “steal” mass from the current chunk. The authors correct that Jensen bias …
A training-free video conditioning trick could matter for anyone building generative tools: it injects low-frequency phase data from a reference clip directly into diffusion noise to steer motion. The appeal …
A new video relighting system can keep full-body human performances temporally stable while changing lighting, which is the hard part most relighting demos still struggle with. BodyReLux is trained on …
A new rigging paper tackles a familiar pain point: auto-rigged characters often have no real pose-space model, so random sampling or manual tweaking can produce broken joints and ugly self-intersections. …
DiffHDR uses video diffusion models to turn ordinary 8-bit LDR footage into HDR-ready video with more usable highlight and shadow detail. The system can also steer the re-exposure with text …
A new physics-informed video diffusion framework offers game developers a faster way to simulate fluid dynamics. By integrating physical constraints directly into the generative process, the method produces realistic water …