Image Generation

Today's Stories

1. 1

   Krea releases open-weights AI image generation models to address concerns that AI-generated visuals look monotonous and lack originality.

   AI creative tools startup Krea opened the weights to its new image models, Krea 2 Raw and Krea 2 Turbo, under a custom license. Both models are available for public download on Hugging Face. Firms with more than 50 seats must pay for Enterprise usage, and all users must implement technical safeguards to prevent generation of illegal materials, non-consensual intimate imagery, child sexual abuse material, or defamatory assets. Enterprises have begun integrating AI-generated images into production workflows, but there is growing concern that AI imagery looks non-distinct, monotonous, and too unoriginal—described as "AI slop." Krea is attempting to address this trend by offering models that provide more visual variety. Open-weights release means developers outside the company can access and build on the technology.

   Both Krea 2 Raw and Krea 2 Turbo are available now on Hugging Face. The custom license structure means adoption hinges on whether enterprises adopt the Enterprise tier and comply with the required safeguards.

2. 2

   Midjourney, an AI image-generation startup, is pivoting into medical imaging with an ultrasound scanner that would scan users in water, but medical experts say the company has provided little evidence to support its claims that the device could match MRI quality.

   Midjourney announced a whole-body ultrasound scanner that would lower users into a vat of water, use a ring of underwater sensors to send sound waves into the body, and generate internal images in about 60 seconds. The company is not initially positioning it as a medical diagnostic device—citing the high cost of FDA clearance and clinical trials—but rather as a wellness tool for spas. Midjourney is moving from its core business of generating synthetic images online into the heavily regulated world of medicine, but radiologists and medical imaging experts told The Verge the company's claims about image resolution and comparisons to MRI are unsupported by evidence. One professor of radiology called the claims "the most grandiose" he has seen, and another noted that ultrasound faces fundamental limits—sound waves cannot pass easily through air or bone—that make it unsuitable for imaging certain body parts.

   The company plans to eventually expand into medical applications, but medical experts emphasized there is "a long road ahead to generating high-quality images and then to understand the clinical value and demonstrate net benefit to patients." Midjourney has shown only low-resolution prototype images so far, and no current evidence exists that detailed ultrasound scans could be comparable to MRI.

3. 3

   AWS is showing how to run ComfyUI—a visual tool for building AI creative workflows—on its SageMaker service to let enterprises generate hundreds of images, videos, or audio at scale in minutes to hours instead of waiting for manual creative work.

   AWS published a technical guide demonstrating how to deploy ComfyUI workflows on SageMaker AI processing jobs, using GPU-accelerated instances and a queue-based architecture that processes multiple requests in parallel. The example uses Z-Image Turbo, a text-to-image model with 6B parameters, to generate batches of high-quality images without manual intervention. For businesses, the speed matters—content delays can mean lost sales and missed marketing deadlines. Automating image, video, and audio generation frees creative teams from repetitive tasks so they can focus on strategy, while the ability to test AI-generated content in controlled environments before global rollout helps protect brand consistency and compliance.

   The solution uses pay-per-second billing with automatic job termination, so enterprises only pay for the compute they actually use. ComfyUI workflows can be exported as JSON and swapped into the deployment, and the architecture scales naturally across thousands of outputs without manual scaling.

4. 4

   Researchers at Shanghai AI Lab introduced Self-Harness, a framework that lets AI agents automatically improve their own operating rules by examining their execution traces, eliminating manual debugging and enabling customization for specific business needs.

   Researchers at the Shanghai Artificial Intelligence Laboratory developed Self-Harness, a system where an LLM-based agent (an AI that makes decisions and completes tasks autonomously) improves its own operating rules by analyzing its own execution traces rather than relying on manual, ad hoc debugging. Most enterprises cannot build their own frontier AI language models, but they can and should customize the harness (the framework controlling how the model operates) for their specific purposes. Today, harness tuning relies heavily on intuition rather than systematic feedback, making it hard to keep pace with evolving models—Self-Harness addresses this by replacing guesswork with empirical evidence.

   Self-improving harnesses could enable development teams to deploy robust custom agents that continually adapt their own execution protocols to overcome model-specific weaknesses, though the article does not specify a release date or availability.

5. 5

   An Inconvenient Truth About AI

   An Inconvenient Truth About AI

6. 6

   Swiss AI Initiative releases Apertus, an open-source foundation model compliant with EU AI Act requirements and competitive with top open models at 8B and 70B parameter scales.

   The Swiss AI Initiative—a collaboration between EPFL, ETH Zurich, and CSCS—has developed Apertus, an open-weights AI foundation model with fully transparent training data, code, weights, methods, and alignment principles. The model is multilingual from day one, trained on 1000+ languages, and meets EU AI Act requirements by respecting opt-outs, removing personally identifiable information, and preventing memorization. Open-source foundation models give organizations and researchers access to AI systems they can audit, modify, and deploy without reliance on proprietary vendors. Apertus's compliance architecture and transparent training suggest a path for building AI that satisfies regulatory obligations while remaining reproducible—a concern for EU-based enterprises and public institutions operating under the AI Act.

   Apertus is positioned as competitive with top open models at equivalent scales of 8B and 70B parameters. Swisscom is a strategic partner of the Swiss AI Initiative, signaling potential industry adoption and real-world deployment pathways.