Matt Wolfe tours voice, video, world-model and 3D research in this 25-minute roundup. The OpenAI Dots segment is sponsored.

Direct the performance

  • Fish Audio Drama 3: write acting directions into the script, add breaths or laughter, and apply effects to selected words.
  • Wolfe distinguishes pronunciation scores from listener preference: clearer words do not automatically produce a favorite voice.
  • OmniHuman 2.0: animate a photo with audio and text directions, at 720p for up to five minutes.
  • BytePlus lists $0.07 per generated second; Wolfe still finds the lip sync unconvincing and flags drifting backgrounds and impersonation risks.

Give generated worlds reliable state

  • WorldCrafter: retrieve earlier views by camera position so revisited places stay more consistent.
  • Matrix-Game 4.0: a game engine owns rules and object positions; a video model repaints its low-detail frames.
  • A coding agent can add rules and update scene prompts, but Wolfe warns about scene-specific models, low resolution and uncertain consumer-GPU performance.
  • StreamRender-H3: NVIDIA gives its renderer a color-coded scene map, separating the world’s behavior from its appearance.
  • NVIDIA reports 218 milliseconds per diffusion-render step on eight GB200 GPUs, with results limited to its racing scenario.

Separate experiments from usable tools

  • ByteDance’s unnamed world-model presentation is event footage without a public release date in Wolfe’s account.
  • Odyssey’s simulated-fly demonstration has no accompanying paper, code or integration explanation in the material he discusses.
  • T3lescope: reconstruct detailed 3D meshes from multiple photographed views through coarse-to-fine refinement.
  • GaussiAnimate: combine an inner skeleton with flexible surface controls for clothing and soft tissue; captured motion sequences are required.
  • WorldGacha: a procedural city generator with image and 3D exports, explicitly a non-AI bonus.

Caveat: this mixes research showcases, vendor demos and Wolfe’s limited hands-on impressions. Availability, licenses and hardware requirements differ substantially.

“The engine handles the gameplay and the AI handles the look.” — Matt Wolfe, 12:03, lightly punctuated