The AI Canvas Newsletter #1

In the pursuit of artificial general intelligence in computer vision, a ground-breaking framework has emerged - InstructDiffusion. This innovative approach transforms diverse computer vision tasks into intuitive image-manipulation processes guided by human instructions. Unlike previous methods that struggled to unify tasks due to varying output formats, InstructDiffusion harnesses the power of denoising diffusion models and treats all tasks as instructional image editing. It excels in handling output formats like RGB images, binary masks, and key points, effectively covering a wide range of vision tasks.

Read More

Step into the world of cutting-edge generative AI with Stability AI's Harmonai lab and their latest creation, "Stable Audio." While diffusion models have revolutionised image and video generation, music has posed unique challenges due to fixed-length constraints. Stable Audio changes the game by allowing audio generation to be guided by text metadata, audio duration, and start time. This approach offers creative control over both content and length, making it perfect for crafting complete songs. With speedy inference and a commitment to open-source accessibility, this 907M parameter U-Net model represents a significant leap in music generation technology.

If you want to learn more, you can read the full summary here.

Microsoft Research continues to push the boundaries of smaller Transformer-based language models in their latest endeavour, "Textbooks Are All You Need II: phi-1.5 technical report." Building on the success of previous models like TinyStories and phi-1, this report explores the concept of common sense reasoning in natural language. Phi-1.5, a 1.3 billion parameter model, outshines models five times its size in natural language tasks, including grade-school mathematics and basic coding.

Find out more here.