FLUX.2
PulseAugur coverage of FLUX.2 — every cluster mentioning FLUX.2 across labs, papers, and developer communities, ranked by signal.
- 2026-05-22 product_launch Black Forest Labs launched the FLUX.2 image generation model. source
13 day(s) with sentiment data
-
New 8-step LoRA for MOSS-Image-Base reduces inference steps
A new 8-step LoRA (Low-Rank Adaptation) model has been developed for MOSS-Image-Base, a diffusion model. This LoRA, trained using a distribution matching technique, significantly reduces the number of inference steps re…
-
ZPix app adds MOSS-Image SFT, SDXL, and FLUX model support
The ZPix application has been updated to include support for the MOSS-Image SFT model, along with several other image generation models including SDXL and FLUX.1 Fill and FLUX.2. This new version also introduces feature…
-
Text-to-image models struggle with object counts beyond ten
A new arXiv paper investigates how text-to-image models handle number words, particularly when requesting more than ten objects. The study found that while models like SD3.5-Medium and FLUX.1-dev struggle to accurately …
-
Open-source tool simplifies AI image prompt generation from booru tags
A developer has created a free, open-source web tool designed to streamline the process of generating prompts for various AI image generation models. The tool converts tags from booru image sites into clean, usable prom…
-
Open-source ComfyUI tools Higgsfield and H3 Continuity Studio released
Two open-source projects, Higgsfield.ai and H3 Continuity Studio, have been released for ComfyUI, offering enhanced capabilities for local AI-powered video and image generation. Higgsfield.ai, a CreatorOS Studio, allows…
-
New distillation method improves AI model training with joint scaling and step-aware critics
Researchers have developed a new method called Distribution Matching Distillation (DMD) to improve the training of few-step student models. By analyzing the critic's capacity and information, they found that scaling the…
-
Morgana 1.2 finetune released for realistic photography generation
Morgana 1.2, a new photography-focused finetune of the FLUX.2 Klein 4B model, has been released. Developed by Kitani, this model aims to generate realistic images with natural lighting, detailed textures, and accurate c…
-
New method improves geometric consistency in AI room defurnishing
Researchers have developed a new training method to improve the geometric consistency of prompt-based image editors when removing furniture from room images. By finetuning a LoRA adaptor on a dataset of furnished and em…
-
New Directional Attention Guidance speeds up diffusion models
Researchers have introduced Directional Attention Guidance (DAG), a novel training-free method to enhance the quality and speed of diffusion models. DAG addresses limitations of Classifier-Free Guidance (CFG) by analyzi…
-
MOSS-Image open model hits 1.3M downloads in four days
MOSS-Image, a 5 billion parameter open image generation model, has seen significant adoption with over 1.3 million downloads in its first four days. Developed by MOSI and the OpenMOSS team, the model is built upon a dif…
-
Visage-Image model released for realistic face generation
A new, small text-to-image model named Visage-Image has been released, specifically designed for generating realistic human faces. The model was trained on a custom dataset and utilizes a single-stream DiT architecture …
-
Unsloth adds Recursive Flint, MOSS-Image, NVIDIA Nemotron Bolt E4B models
Unsloth has released version 0.1.910-beta, introducing support for several new models including Recursive's Flint and MOSS-Image, as well as NVIDIA's Nemotron Bolt E4B. This update enables users to run these models loca…
-
AI Image Generation Models Exhibit Peculiar Behaviors, Defying Prompts
An article explores the inner workings of AI image generation models, specifically diffusion models, latent space, and flow matching. It highlights peculiar behaviors observed during image generation, such as AI clocks …
-
New method unlocks precise spatial control in image synthesis models
Researchers have identified a critical factor for improving spatial control in flow-matching models used for image synthesis: the signal-to-noise ratio (SNR) distribution during training. By shifting focus to high-noise…
-
Hugging Face hosts lodestones/Kromoss image generation model
The lodestones/Kromoss model, a fine-tune of MOSS-Image-Base, has been released on Hugging Face. This model aims to improve upon the base model's blurriness and offers compatibility with various libraries and applicatio…
-
4LC introduces 4D latent cache for efficient video generation · 2 sources tracked
Researchers have introduced 4LC, a novel 4D latent cache designed to improve efficiency in dynamic world modeling for video generation. This system reuses static and dynamic visual content by binding representations fro…
-
OpenMOSS releases MOSS-Image text-to-image models
OpenMOSS has released two text-to-image models, MOSS-Image-Base and MOSS-Image-SFT, both derived from the MOSS-Image family. These models are designed to support natural language prompts, tag sequences, and JSON layouts…
-
New research reveals aligned color subspace in VAEs for text-to-image models
Researchers have identified a consistent color subspace within the latent spaces of Variational Autoencoders (VAEs) used in text-to-image models. This subspace aligns with brightness and opponent-color axes, a finding d…
-
Opus 5.5 used to build Wikipedia interface replacement
A user on Reddit shared their experience using Opus 5.5 to create a replacement interface for Wikipedia. The interface, named folio.helikos.dev, utilizes Gemini for page content, FLUX.1 on Cloudflare Workers AI for main…
-
New AI methods enhance image editing by verifying edits and reducing unintended changes · 2 sources tracked
Researchers have developed new methods for improving instruction-guided image editing models. One approach, ARRO, uses a reinforcement learning agent with a vision-language reward model to audit edits for both requested…