New speech and video models expand creative controls

Fish Audio’s Drama 3 is a 36-billion-parameter open-weight text-to-speech model that accepts plain-English acting directions rather than emotion presets. It typically requires about 46GB of VRAM, while a community-made 4-bit quantized version runs on roughly 32GB. Neither configuration fits on most laptops. API pricing starts at about ? per million characters at launch, rising to ? after November 12th.

ByteDance’s Omnihuman 2.0 turns a single image and audio track into continuous talking or singing videos lasting up to five minutes, compared with 60 seconds previously. It renders at 720p and accepts text prompts for camera movements. Pricing has fallen from 12 cents per second to 7 cents per second.

Tencent’s Worldcrafter addresses world-model memory loss through camera-position lookup. Skyworks’ Matrix Game 4.0 instead uses a traditional game engine to maintain positions and rules while a video model repaints low-poly frames, preventing off-screen objects from disappearing. Each scene uses its own roughly 5-billion-parameter model and runs at 480×864, with weights and inference code available under Apache 2.0. Nvidia’s Stream Render H3 repaints a flat, color-coded map in real time using data-center GPUs.