PEVA: Whole-Body Motion Guided Ego-Centric Video Prediction

Researchers from BAIR (Berkeley AI) introduced the PEVA (Predicting Ego-centric Video from human Actions) model, which generates next first-person video frames based on past frames and a sequence of human actions specified via 3D pose changes. The model uses an autoregressive conditional diffusion transformer trained on the Nymeria dataset and can predict atomic actions, simulate counterfactual scenarios, and support long video generation up to 16 seconds. PEVA can also be used for planning by evaluating different action sequences based on similarity to a target image.

Berkeley Artificial Intelligence Research (BAIR)

Mistral AI unveils Devstral 2 family of code models and Mistral Vibe CLI tool

Mistral AI has released the next generation of programming models, Devstral 2, in two sizes (123B and 24B parameters), available under open licenses. The models demonstrate high performance on the SWE-bench Verified benchmark despite smaller size and cost. Also introduced is the CLI agent Mistral Vibe for autonomous development task solving.

Mistral Mistral AI