Anthropic released Claude Fable 5.2 and Claude Mythos 5.2, with Fable 5.2 demonstrating an image-making path that bypasses image generation models entirely: it writes 1,000-2,000 lines of Python to drive a simulated painting engine, controlling brush texture, paint wetness and canvas grain stroke by stroke. In a staff demo, the model produced a parking-lot shopping cart using roughly 700,000 simulated brushstrokes, with no external image model or pre-rendered base image.
Diffusion Model
扩散模型
Generative modeling approach that produces images by iteratively denoising random noise, underlying mainstream text-to-image too
Curated Intelligence
MBZUAI's Institute of Foundation Models released PAN 2, an open foundation world model for physical intelligence built around simulative reasoning that lets agents play out outcomes before acting. It splits reasoning into two tracks: an LLM determines what happens, when, and how the camera moves and writes it as text, while a 29B diffusion model trained from scratch renders that text into visuals. Ablating the LLM collapses multi-step performance to 8.7% correct steps with no completed plans, whereas the combined system reaches 75.1% correct steps and 34.0% fully completed plans.
Video Rebirth released HyperFlow H3 Flash, a compressed few-step variant of MiniMax-H3 that shrinks a 31B joint video-audio diffusion model to 13.7B while supporting 4/6/8-step sampling without CFG. Weights, ComfyUI nodes and GGUF deployment are open-sourced; a 5.2s clip renders in ~92s on one RTX 5090 and 8.7s on 4×B200, with real-time inference on 4 Blackwell GPUs via SOL-H3.
OpenAI introduces Images 2.6 across ChatGPT, Codex, and the API, adding GIF animation support and improved editing consistency. The update boosts generation speed by 50% and offers developers two distinct models: Flare for low-latency tasks and Sunburst for high-fidelity visuals.
Startup Dalet Labs raised a $7M pre-seed led by Lightspeed to build an AI-generated shopping feed. The system uses LLMs to extract attributes from liked images for diffusion prompts, reducing hallucinations and enabling self-growing content without photoshoots.
This article offers a comprehensive guide to AI image generation, covering diffusion model principles, historical evolution, and the practical performance of tools like ChatGPT. Based on experiments calling multiple image providers via Python, the author analyzes pros and cons of different models and discusses integrating image generation into code and agents.
The University of Washington and NVIDIA team published RFdiffusion3 in Science, enabling atomic-level modeling of amino acid coordinates to design functional proteins. Experimental validation confirms that enzymes designed by the system exhibit expected catalytic activity in real reactions, with generation speed approximately ten times faster than previous versions.
Envato integrates an 'Edit with AI' feature into its asset library, enabling localized modifications on high-quality real stock photos. By preserving the original subject, framing, and lighting while only generating selected areas, this approach aims to mitigate common structural failures like malformed limbs and physical logic errors found in pure text-to-image models.
Shengshu Tech released the Vidu Q4 Preview, enhancing character expressiveness and multi-style adaptability while cutting generation costs to under $0.01 per second. This update aims to empower ordinary creators with affordable iteration budgets, shifting AI video from tech demos to practical production.
Researchers from UPenn and UIUC introduce Normalizing Trajectory Models (NTM), which use conditional normalizing flows to retain exact likelihood in few-step diffusion sampling. By combining shallow invertible blocks with a deep parallel predictor, NTM achieves state-of-the-art text-to-image results in just four steps while enabling self-distillation through its precise trajectory likelihood.