A rapid-fire look at all the AI updates to close out 2025:
Video & World Models:
Kling 2.6 Motion Control - Kling AI feature enabling precise, reference-image–driven control of character actions, expressions, and motion for up to 30 seconds.
LongVie2 - End-to-end autoregressive video world model generating up to 5 minutes with strong controllability and temporal consistency
Lucy Motion - Decart's image-to-video with path-based motion control where you draw trajectories and the model animates exactly along them
Odyssey-2 Pro - Real-time world model that generates every pixel frame-by-frame, simulating physics forward for minutes
Wan Vision Enhancer - fal's in-house video upscaler with 5 creativity levels from faithful restoration to artistic reimagining
DAAAM - Builds hierarchical 4D scene graphs as spatio-temporal memory, enabling agents to describe anything, anywhere, at any moment
Text-to-Image:
FLUX.2 [max] - Black Forest Labs' top-tier quality with grounded web search, up to 10 reference images for consistency, ranked #2 on Artificial Analysis
GPT Image 1.5 - OpenAI's API release with improved instruction following, better text rendering, and enhanced preservation of logos and faces
Image-to-3D:
Qwen-Image-Layered - Alibaba’s AI model that decomposes images into editable RGBA layers without external segmentation
SHARP - Apple’s feedforward system that converts a single 2D image into a photorealistic 3D scene in under a second, supporting real-time rendering at 100+ FPS
TRELLIS.2 - Microsoft's open-source 4B-parameter model producing up to 1536³ PBR textured assets with arbitrary topology support
Aurora - Creatify's image-to-avatar model creating ultra-realistic avatars from a single image with expression control and scalable quality
Audio Separation:
SAM Audio - Meta's first unified model isolating any sound from complex audio mixtures using text, visual, or span prompts
Surprise Entry in the AI Game:
Zoom AI Companion 3.0 - Achieved 48.1% on Humanity’s Last Exam—an exceptionally rigorous benchmark for expert-level knowledge and deep reasoning—marking a 2.3% gain over the previous state of the art through improved summaries and reasoning.
