Today in Generative Media

Google's Veo 2 genAI video model; AI studio Promise raises seed; TIME's genAI

News and Opinion

• Google DeepMind unveils a new video model to rival Sora (TechCrunch)

  • Veo 2: Our state-of-the-art video generation model (Google DeepMind)

• Generative AI in Hollywood: Promise Studio Raises Seed Money to Innovate Filmmaking (eWeek)

• Why We’re Introducing Generative AI to TIME’s Journalism (Time)

• A ‘Newspeak’ world - if generative AI will be used to manipulate, who stands to benefit? (Diginomica)

• The future of content creation and production with generative AI (Adobe blog)

• Meta asks the government to block OpenAI’s switch to a for-profit (The Verge)

• Generative AI’s Continuing Copyright Problems, an Essay in Memory of Suchir Balaji, 1998 - 2024 (Gary Marcus on Substack)

• Behind the scenes: How Paramount+ used Adobe Firefly generative AI in a social media campaign for the movie IF (Adobe blog)

Software

• Gemini 2.0 Flash ushers in a new era of real-time multimodal AI (VentureBeat)

• Microsoft releases Phi-4, a 14-billion parameter model which performs at par with GPT-4o-mini and recently released Llama-3.3-70B (X)

• genmoai / mochi-1-lora-trainer (Replicate)

• fofr / flux-condensation: Flux fine-tuned to write and draw in condensation (replicate.com)

• DOGS: Distributed-Oriented Gaussian Splatting for Large-Scale 3D Reconstruction Via Gaussian Consensus (GitHub)

Research

• Mobile Video Diffusion (project page)

• Neural Localizer Fields for Continuous 3D Human Pose and Shape Estimation (project page)

• Neural LightRig: Unlocking Accurate Object Normal and Material Estimation with Multi-Light Diffusion (project page)

• PrEditor3D: Fast and Precise 3D Shape Editing (project page)

• LLaMA-Mesh: Unifying 3D Mesh Generation with Language Models (project page)

• Feat2GS Probing Visual Foundation Models with Gaussian Splatting (project page)

• Coherent 3D Portrait Video Reconstruction via Triplane Fusion (project page)

• PlanarSplatting: Accurate Planar Surface Reconstruction in 3 Minutes (project page)

• Spectral Image Tokenizer (arXiv)

• SLAM3R: Real-Time Dense Scene Reconstruction from Monocular RGB Videos (arXiv)

Misc

• I visualized a dozen Gemma SAE latents and for each I generated a 144 megapixel (9k x 16k) image over 4000 unique activating prompts. Here's what it looks like zooming in twice to highlight some of the patterns that emerge. (X)