FLUX.2 - An open-source AI image generation and editing model from Black Forest Labs.

What is FLUX.2?

FLUX.2 is an AI image model developed by Black Forest Labs, designed specifically for real-world creative workflows. The model supports multi-image references of up to 10 images, generating high-quality images up to 4MP resolution with exceptional detail and text rendering capabilities. FLUX.2 is available in multiple versions, including the high-performance FLUX.2 [pro], the customizable FLUX.2 [flex], the open-source FLUX.2 [dev], and the upcoming FLUX.2 [klein]. Combining a visual language model with a stream transformer architecture, the model significantly improves real-world knowledge understanding and image generation quality, driving open innovation and widespread application of visual intelligence technologies.

Main features of FLUX.2

• Multiple images for referenceThe model supports referencing up to 10 images simultaneously, maintaining consistency in characters, style, and products.
• High-resolution image generationThe model supports image editing up to 4MP, making it suitable for product photography, visualization, and photographic applications.
• Complex text renderingThe model can handle complex layouts, infographics, emojis, and UI designs, and supports readable small text.
• Instruction compliance capabilityImproved adherence to complex, structured instructions, including multi-part hints and combined constraints.
• Real-world knowledgeIt performs better in terms of lighting, spatial logic, and scene coherence, generating more realistic images.

Technical Principles of FLUX.2

• Latent Flow Matching ArchitectureFLUX.2 employs a latent flow matching architecture. By performing flow matching in the latent space, the model can efficiently handle image generation and editing tasks while maintaining the coherence and consistency of the generated images. This architecture makes FLUX.2 perform exceptionally well in handling complex image synthesis tasks, especially in multi-image reference and high-resolution generation.
• Coupling of visual language model and stream converterFLUX.2 combines a Mistral-3 24B parameter Visual Language Model (VLM) with a Transformer. The VLM provides the model with rich real-world knowledge and semantic understanding, enabling FLUX.2 to better understand complex cue words and scene logic. The Transformer focuses on capturing spatial relationships, material properties, and compositional logic in images, compensating for the shortcomings of traditional architectures. This coupling makes FLUX.2 excel in generating complex scenes and details, especially when handling multi-image references and complex text rendering.
• Optimization of Variational Autoencoder (VAE)FLUX.2 introduces a novel variational autoencoder (VAE) for optimizing latent representations. The VAE offers an optimal trade-off between learnability, image quality, and compression ratio. By retraining the latent space, FLUX.2 addresses the learnability-quality-compression trilemma, achieving higher image quality and better generation efficiency.
• Multiple images for reference and style consistencyFLUX.2 supports referencing up to 10 images simultaneously, using advanced multi-image fusion algorithms to ensure consistency in style, character, and product details in the generated images. This multi-image referencing capability makes FLUX.2 particularly suitable for creative workflows that require maintaining brand style or scene coherence, such as advertising design, product visualization, and film post-production.

FLUX.2 project address

• Project official websitehttps://bfl.ai/blog/flux-2
• HuggingFace model libraryhttps://huggingface.co/collections/black-forest-labs/flux2

How to use FLUX.2

• FLUX.2 [pro]It can be used directly through BFL Playground or BFL API, making it suitable for production environments without the need for local deployment.
• FLUX.2 [flex]It can be used via bfl.ai/play or the BFL API, and the generation parameters can be adjusted, making it suitable for developers who need fine control.
• FLUX.2 [dev]Access the Hugging Face model library, download the open weight model, and run it locally with the reference inference code. This is suitable for developers to perform customized development.
• FLUX.2 [klein](Coming Soon): The open-source version of FLUX.2 is suitable for developers to participate in Beta testing (https://docs.google.com/forms/d/e/1FAIpQLScOIvOkHN2fPbD8cFsAf7MQJfqu2bnEmoNb0x1k3ismTLLm-Q/viewform) for local experimentation and innovation.
• FLUX.2 – VAEA novel variational autoencoder for latent representations, which serves as a foundational component supporting other FLUX.2 models and can be used with the Hugging Face model library.

Application scenarios of FLUX.2

• Advertising productionFLUX.2 can quickly generate high-quality product advertising images, supports multiple image references to maintain brand style consistency, and can generate creative advertising content based on complex prompts.
• UI/UX DesignThe model supports complex layouts and text rendering, and can generate user interface prototypes and design drafts to help designers quickly realize their creative ideas.
• Brand promotionCreate visual content for brands through high-resolution image generation and editing, ensuring brand consistency across different media.

• Film and television special effectsIt is used to generate realistic scenes, props and characters, and supports multiple image references to maintain the consistency of visual style, reducing the time and cost of special effects production.
• Animation ProductionIt accelerates the animation production process while maintaining consistency in animation style by generating high-quality animation frames and backgrounds.