OpenAI Whisper, Stable Diffusion inpainting/outpainting, DALL E API? - AI NEWS

OpenAI’s Whisper is the week’s biggest practical upgrade: an open-source speech-recognition neural network aimed at near-human robustness and accuracy for English. The pitch is straightforward—when phones mishear speech, Whisper is designed to do better, including with fast delivery and heavy accents. Examples in the demo include speed-talking and a thick-accent passage that would typically be hard for consumer transcription systems; the transcripts come out clean enough to suggest the model is handling difficult audio conditions more reliably than standard mobile transcription.

Whisper also matters because it’s not just a “transcribe English” tool. The workflow described includes handling speech from other languages and converting it into English text, which points toward real-time, cross-language communication—someone speaks, the system transcribes and translates automatically, and the other person reads it in their own language. The transcript walkthrough touches on how the token-based transcription process works and notes that OpenAI provides a GitHub/Collab example for trying it out, making it easier for developers to integrate Whisper into phones and apps.

The second major development is Dream Studio’s new editing system for Stable Diffusion, positioned as a full inpainting and outpainting editor rather than a basic image-to-image variation tool. The interface adds controls like select-and-move for repositioning elements, scaling for resizing, and a masking brush for targeted edits. Compared with earlier DALL·E-style editing workflows, the editor adds more granular controls such as blur size, strength, and image opacity—useful for blending edited regions back into the original image. A “restore” option is also highlighted, letting users recover parts of the original image after erasing.

A hands-on example centers on adding “yellow legs” to a lemon character. The editor blurs around the edited area while generating, then produces results that actually add the requested legs. The workflow supports multiple generations (including up to nine variations) so users can iterate quickly, and the interface includes advanced settings such as CFG scale and sampler choices, with the ability to apply different settings across parts of an image. An image-to-image mode is also available, where a prompt like “a photo of a lemon” shifts the style toward more photorealistic outputs, with strength settings controlling how closely the result tracks the original.

Beyond these two concrete releases, the transcript flags rumors and ecosystem momentum: a possible Midjourney app appearing on the Apple App Store (with a stated expected date), a surge in third-party AI image generator apps across app stores, and chatter that OpenAI may allow other companies to purchase access to the DALL·E API. Taken together, the theme is clear—speech transcription is getting more accurate and more accessible via open tooling, while image generation is moving from “generate once” toward interactive, controllable editing inside dedicated web and app-style interfaces.