Descriptions:
This weekly AI news roundup from AI Search opens with what may be its most technically striking item: StreamDiff VSR, an open-source real-time video upscaler benchmarked at 0.3 seconds per frame on a single RTX 4090, compared to over 4,600 seconds for standard diffusion-based video super-resolution. The demo shows visibly sharper output on noisy, blurry, and shaky source footage, and the GitHub repository is already public with local setup instructions.
Other major releases covered include UltraShape 1.0, a new open-source 3D model generator that outperforms Trellis 2 on complex subjects like detailed characters and organic forms — though it produces geometry only, without surface texture. Ume 1.5 is highlighted as a unified image and video editing model supporting natural language instructions for both modalities, demonstrated by tasks like adding accessories to animals in video while preserving motion dynamics. DeepSeek’s return with a significant new release is flagged as a headline story for the week.
Additional tools include a real-time interactive 3D world generator, Qwen ImageEdit 2512 for 1–2 second offline image generation, an AI camera-angle manipulation system for existing video, and Tencent’s 3D character animation pipeline. MiniMax M2.1, an open-source language model claiming top results on multi-SWE-Bench and multilingual coding benchmarks — including outperforming Gemini 3 Pro and Claude 4.5 — is also featured. All projects link to public GitHub repositories.
📺 Source: AI Search · Published January 04, 2026
🏷️ Format: Roundup
