Tag
Meta introduced Muse Image and previewed Muse Video, an agentic image and video generation system that enables precise edits, multiple references, and integration with Instagram context, turning media generation into a full creative operating system.
M87 is an early-preview text-to-image AI model specialized in generating analog film-style photographs, available on Hugging Face.
This paper proposes a generator-agnostic post-generation curation method that selects informative subsets of synthetic images by splitting real classes into canonical homogeneous and non-redundant heterogeneous subsets, and scoring synthetic images via a fidelity-diversity criterion. It consistently outperforms existing data-selection baselines and matches real-data performance with up to 40% fewer synthetic samples.
This post details the author's setup and benchmarks for running NVFP4-quantized image and video generation models on a GIGABYTE AI TOP ATOM (DGX Spark) workstation, achieving impressive performance with models like FLUX.2, Qwen-Image, and LTX-2.3 for video with synchronized audio.
Muse Image and Muse Video are new AI models for image and video generation, offering precise instruction following, editing, multi-reference composition, and social context from Instagram, along with high visual fidelity and native audio support.
Midjourney is seeking to compel Disney, Universal, and Warner Bros. to disclose details of their own AI usage as part of a copyright infringement lawsuit, arguing that the studios' internal AI practices could support its fair use defense.
The author shares their experience building a small flow matching image generation model trained on Apple emoji images, describing the initial failed approach and the successful pivot using RGB channels, residual blocks, and attention.
A developer launches crdible, a tool that takes rough ideas or voice notes and generates LinkedIn posts in the user's voice along with non-AI-looking images.
Claude Fable 5 is back online, and the prompts it writes can make Grok generate videos comparable to Seedance 2.5 in quality and feel at a 6x lower cost, with detailed portrait prompt examples.
Boogu-Image-0.1 is a new unified image generation and editing model family with 10B parameters, available under Apache 2.0 license. It features fast turbo inference in 4 steps, trained on 10x less data, and supports Chinese and English.
This model introduces a depth-conditioned ControlNet-LoRA for Krea-2, enabling depth-map-guided image generation with high depth consistency (0.98-0.99 Pearson correlation). It supports both Raw and Turbo variants and includes easy inference scripts and Comfy UI integration.
Diffusers library has been updated with new image and video pipelines including Ideogram4, MotifVideo, and the DiffusionGemma model.
Speculation about a potential stealth model drop from OpenAI on arena.ai called kyros-alpha, which generates images that pass OpenAI's Verify detection but shows some atypical traits.
ComfyClaw is an agentic skill evolution framework for ComfyUI image generation workflows, using typed graph editing and region-level VLM verifiers to translate visual failures into repair suggestions, outperforming baselines across multiple configurations.
Perceptual Flow Matching supervises flow matching in perceptual feature space, enabling high-quality few-step generation with 4-8 sampling steps instead of 35-50, without needing teacher models.
Modular's MAX inference serving achieves 3x faster image generation for FLUX.2-dev than competitors, as per Artificial Analysis benchmarks.
OrbitQuant introduces a data-agnostic quantization method for diffusion transformers that eliminates the need for recalibration across timesteps and modalities, achieving state-of-the-art post-training quantization at low-bit settings for models like FLUX.1 and CogVideoX.
This paper investigates the mechanisms behind self-alignment methods in diffusion transformers, revealing that performance improvements from methods like Self-Flow primarily come from data augmentation along the noise dimension rather than token interactions between noise levels. The authors introduce Attention Separation to demonstrate this and propose an effective design combining self-representation alignment with dual-timestep augmentation.
This paper introduces Representation Distribution Matching (RDM), a method for one-step image generation by matching feature distributions under pretrained encoders, achieving state-of-the-art results on ImageNet and enabling post-training of FLUX.2 into a one-step generator with improved performance.
OTCache is a training-free framework that uses optimal transport to predict caching schedules for diffusion models, achieving up to 4.7x acceleration on FLUX.1, Qwen-Image, and HunyuanVideo while improving generation fidelity.