This weekend in Generative Media

ChatGPT trims safety; Sora biases; Bloomberg's rocky AI start; AI fantasy wavers

News and Opinion

• OpenAI peels back ChatGPT’s safeguards around image creation (TechCrunch)

• OpenAI’s Sora Is Plagued by Sexist, Racist, and Ableist Biases (Wired)

• Bloomberg Has a Rocky Start With A.I. Summaries (New York Times)

• Google is rolling out Gemini’s real-time AI video features (The Verge)

• The Tech Fantasy That Powers A.I. Is Running on Fumes (New York Times)

Software

• TripoSG: High-Fidelity 3D Shape Synthesis using Large-Scale Rectified Flow Models (GitHub)

• TripoSF: High-Resolution and Arbitrary-Topology 3D Shape Modeling with SparseFlex (GitHub)

• we improved our already fastest Wan2.1 image-to-video at FAL and here are its latest stats: > 480P output video generation time: 20 seconds > 720P output video generation time: 40 seconds (X)

Research

• Hi👋3DGen: High-fidelity 3D Geometry Generation from Images via Normal Bridging (project page) Demo on HuggingFace.

• VideoMind: A Chain-of-LoRA Agent for Long Video Reasoning (project page)

• LeX-Art: Rethinking Text Generation via Scalable High-Quality Data Synthesis (project page)

• Visual Jenga: Discovering Object Dependencies via Counterfactual Inpainting (project page)

• Synthetic Video Enhances Physical Fidelity in Video Synthesis (project page)

• Can Video Diffusion Model Reconstruct 4D Geometry? (arXiv)

• GenFusion: Closing the Loop between Reconstruction and Generation via Videos (arXiv)

Misc

• Art is now accessible. (X)

• Did a new training run for this timelapse growth lora and it's looking beautiful (X)

• The new ChatGPT Image tool has upped the AI storytelling game, allowing you to block out an entire scene with consistent characters and camera angles. Watch for yourself. (X)