computer vision

Mistral AI reports that fine-tuning its Pixtral-12B vision model on satellite imagery increased classification accuracy from 56% to 91% on the Aerial Image Dataset. The company used LoRA (Low-Rank Adaptation) to train on 8,000 samples for under $10, reducing hallucinations from 5% to 0.1%.

Google Photos announced an AI feature that automatically creates a digital wardrobe from clothing photos in users' libraries. The feature allows outfit mixing and virtual try-on, launching on Android this summer before expanding to iOS.

OpenAI launched ChatGPT Images 2.0, upgrading from decorative images to full-page graphics with detailed text rendering. The update is available to all ChatGPT tiers, with advanced features requiring paid subscriptions that access the Thinking model. Hands-on testing shows significant improvements in text accuracy and brand-style replication, though factual errors still occur.

Baidu has released ERNIE-Image, an 8B parameter text-to-image generation model built on a single-stream Diffusion Transformer architecture. The model is designed for complex instruction following, text rendering, and structured image generation, and can run on consumer GPUs with 24GB VRAM.