image-generation

Elon Musk announces that Grok Imagine can generate beautiful fashion images.

This paper introduces Type-Aware Repair Allocation (TARA), a training-free framework that decomposes text-to-image prompt optimization into atomic repair allocation, where each failed proposition is routed to a type-conditioned repair operator. Experiments show TARA achieves the best semantic accuracy on DSG and TIFA benchmarks across four generators, improving over VisualPrompter while maintaining image quality.

Qwen-Image-3.0 is a third-generation foundational image generation model supporting up to 4.5k token input, native rendering in 12 languages, and simulation of interfaces like web pages and games, leveraging rich world knowledge for practical deployment.

An experiment pits four vision models (GPT-5.6 Sol, Claude Fable 5, Grok 4.5, Gemini 3.6 Flash) against each other in a colored-pencil drawing arena, tasking them with reproducing targets like the Mona Lisa and Starry Night via tool use. The results reveal surprising cost and quality trade-offs, with Grok struggling and Claude underperforming despite higher cost.

Qwen Image 3, a new image generation model, announced. It generates images in a single pass and could spawn edtech/industrial training startups.

Qwen-Image-3.0 is a new image generation model from Alibaba, offering rich content, authentic details, and deep knowledge capabilities.

Vaibhav Srivastav used ChatGPT with memory to build a personal exercise tracking website, leveraging image generation, a database, and ChatGPT authentication.

Mage-Flow is a compact 4B-parameter generative stack for efficient text-to-image generation and instruction-based image editing, featuring a co-designed lightweight tokenizer (Mage-VAE) and a native-resolution multimodal diffusion transformer trained with rectified flow matching. It achieves competitive performance while enabling high-resolution generation at 0.59s on a single A100 GPU.

Introduces appearance pointers, compact tokens that guide Diffusion Transformers to apply correct appearance cues at specified spatial locations, enabling modality-agnostic localized multimodal control without retraining the base model.

gimi-illustration-skill 2.0 released, supporting custom IP characters. Upload a front-view image to consistently use the same character in later illustrations, eliminating the need for repeated descriptions. An open-source tool for graphic creators, professional writers, etc.

Introduces Three-Body Scattering Modeling (TBSM), a framework for one-step generative modeling that learns a transport field. Achieves competitive results on ImageNet-256 and scales to text-to-image models with up to 20B parameters.

FlowMimic presents a method for mask-free visual editing and generation across video and image modalities using pixel-pair warped flow fields, enabling real-time video editing data generation from image editing samples and aligning modality capabilities through mimicry losses.

Proposes DiFA, a training-free framework that reframes inference-time data prediction refinement as sequential state estimation using Kalman filtering, significantly improving generative fidelity on CIFAR-10 and ImageNet.

Demonstrates deep dream effects on modern LLMs by optimizing images to maximize probability of a target caption; Gemma 12B (no vision encoder) stamps recognizable objects, while E4B (with vision encoder) drifts to texture.

Introduces Self-Correcting Coupled Markov Jump Processes (SC-CMJP) and a training-free sampler CO2Jump for concurrent image understanding and generation, achieving state-of-the-art joint performance on editing, maze, and nonogram tasks.

MeanFlowNFT introduces a forward-process reinforcement learning method for average-velocity generators, enabling efficient alignment with human preferences while preserving fast few-step sampling. Experiments show it outperforms prior RL-tuned few-step generators on most metrics and even surpasses multi-step RL-tuned diffusion models.

Simon Willison creates a custom 'pet' for Codex Desktop — a pelican on a bicycle — entirely via AI image generation using GPT-5.6 Sol and gpt-image-2, documenting the open-source process and sprite generation.

Google marks 25 years of Google Images by introducing a new browsable image gallery and AI-powered image generation in Search using the Nano Banana model, enabling users to create custom visuals from text prompts.

Google is redesigning the Google Images homepage to show a dynamic, personalized gallery of images before you search, and will soon allow AI Overviews to generate images using its Nano Banana 2 Lite model.

Spatially Speculative Decoding (SSD) accelerates autoregressive image models by predicting entire rows in parallel using small helper networks, achieving up to 13.28x speedup while maintaining benchmark performance.