Daily Digest — 2026-08-30
4 items · 1 research labs, 3 industry media
🏛️ Research Labs (1)
Celebrating 25 years of visual search innovation
Google Images celebrates its 25th anniversary with two major updates: a dynamic, immersive gallery for browsing images tailored to user interests, and image generation directly within AI Overviews using the Nano Banana model. These innovations build on historical milestones such as Similar Images (2009), Search by Image (2011), Google Lens (2018), Multisearch (2022), and Circle to Search (2024). Recent advancements include Lens + AI Mode (2025) for multimodal reasoning, Search Live (2025) for interactive video-based queries, and Circle to Search Multi-Object Recognition (2026) for simultaneous exploration of multiple objects in an image. These developments collectively enhance visual search capabilities across devices.
multimodal searchimage generationvisual image fan-outai modecircle to search
📜 arXiv Papers
No new items today.
📰 Industry Media (3)
Building Custom Batched Ensemble Weather Forecasting with NVIDIA Earth2Studio
The tutorial demonstrates a custom batched ensemble weather forecasting workflow using NVIDIA Earth2Studio, integrating a Fully Convolutional Network (FCN) prognostic model and a wind-power diagnostic. Initial conditions are retrieved from the Global Forecast System (GFS), and perturbations are applied with variable-specific noise amplitudes. The pipeline executes ensemble forecasts, computes latitude-weighted RMSE, fair CRPS, and spread-skill ratios, and visualizes ensemble uncertainty through spatial maps and spaghetti plots. Results show effective verification against GFS analyses, with metrics quantifying forecast skill and ensemble spread.
ensemble forecastingfully convolutional networklatitude-weighted rmsefair crpsspread-skill ratio
Google AI Releases Gemini Omni 1.1 Flash: 40-Second Scene Extension, First/Last Frame Control, and 4K Upscaling
Google AI introduced Gemini Omni 1.1 Flash, a multimodal video generation and editing model with enhanced controllability. Key updates include scene extension leveraging 10 seconds of prior context (extendable to 40 seconds), first/last frame interpolation for camera control, and 360p-to-4K resolution scaling. The model processes text, image, audio, and video natively via the Interactions API, preserving state through previous_interaction_id. Outputs are billed at ~$0.10 per second of 720p video (5,792 tokens/sec) with SynthID watermarking. Early adopters include Adobe, Figma Weave, and Runway. Limitations include no mid-clip insertion, unsupported audio references, and English-only evaluation.
multimodal video generationscene extensionframe interpolationsynthid watermarkinginteractions api
Hugging Face Unveils Microduck: A $399 Open-Source 25 cm Biped You Train with Reinforcement Learning
Hugging Face and Pollen Robotics introduce Microduck, a $399 open-source bipedal robot trained via reinforcement learning. The 25 cm robot features 15 motors, a Rockchip RK3566 processor, LiDAR, dual IMUs, and a 1-hour battery. Policies are trained in microduck_rl using MuJoCo Warp and PPO, achieving usable gaits in 1–2 hours on CUDA GPUs with 4096 parallel environments. Sim-to-real transfer incorporates domain randomization for voltage, friction, and ±1° gear backlash. Pre-trained behaviors include walking, self-recovery, and roller-skating. Training environments and reward functions are publicly available on GitHub, though mechanical and electronic designs remain proprietary.
reinforcement learningsim-to-realdomain randomizationproprioceptionparallel environments
Generated automatically at 2026-08-29 21:31 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.