FLUX.1 [dev]

FLUX.1 [dev] is an open-weight text-to-image model that represents a significant advancement in image synthesis technology. Built on a hybrid architecture combining multimodal and parallel diffusion transformer blocks, the model is scaled to 12 billion parameters. It employs sophisticated techniques including flow matching, rotary positional embeddings, and parallel attention layers to enhance both performance and hardware efficiency.

The model utilizes guidance distillation during training to improve efficiency while maintaining high-quality output. This approach allows FLUX.1 [dev] to achieve performance levels that surpass many prominent models in the field, including Midjourney v6.0, DALL·E 3 (HD), and SD3-Ultra, particularly in areas such as visual quality, prompt adherence, size/aspect variability, typography, and output diversity.

Capabilities and Performance

FLUX.1 [dev] demonstrates exceptional versatility in image generation, supporting various aspect ratios and resolutions up to 2.0 megapixels. The model has been specifically fine-tuned to preserve output diversity from pretraining, ensuring a wide range of creative possibilities.

It's important to note that while the model excels in many areas, it may occasionally fail to generate outputs perfectly matching prompts, with prompt following significantly influenced by prompting style. The model is not designed to provide factual information and may reflect existing societal biases.

Model Family and Variants

FLUX.1 [dev] is part of a three-variant model family, each designed for specific use cases:

1. FLUX.1 [pro]: The highest-performing variant in the family
2. FLUX.1 [dev]: The open-weight version for non-commercial use
3. FLUX.1 [schnell]: A faster model optimized for local development and personal use, released under Apache2.0 license

Licensing and Usage Considerations

FLUX.1 [dev] is released under a non-commercial license that permits personal and scientific use. The license explicitly prohibits several uses, including:

• Violations of applicable laws
• Exploitation of minors
• Generation of harmful misinformation
• Dissemination of personally identifiable information for malicious purposes
• Harassment of individuals
• Creation of non-consensual nudity or illegal pornography
• Automated decisions affecting legal rights
• Large-scale disinformation campaigns

References

• Black Forest Labs Announcement - Official announcement and detailed model information
• FLUX.1 [dev] GitHub Repository - Reference implementation and sampling code
• FLUX.1 [dev] Non-Commercial License - Detailed licensing information
• FLUX.1 [dev] Acceptable Use Policy - Usage guidelines and restrictions
• ComfyUI GitHub Repository - Local inference implementation
• Diffusers Documentation - Integration guide for the diffusers library