FLUX.2 [dev] Flash is a high-speed, distilled version of the FLUX.2 [dev] model, developed by Fal in collaboration with Black Forest Labs. It is engineered to deliver high-fidelity image generation with significantly reduced latency, typically achieving sub-second inference times on modern hardware. This model leverages advanced distillation techniques to maintain the capabilities of the original 32-billion parameter architecture while requiring only 4 to 8 sampling steps.
The model's architecture combines a Rectified Flow Transformer with the Mistral-3 24B vision-language model (VLM), providing it with sophisticated contextual understanding and spatial reasoning. This synergy enables native support for high-resolution outputs up to 4 megapixels and superior typography rendering. Additionally, it utilizes an improved FLUX.2 VAE designed for enhanced texture detail and color accuracy compared to its predecessors.
Key Capabilities
FLUX.2 [dev] Flash introduces advanced control features, including native support for JSON-structured prompts, which allows for granular specification of scene composition, object attributes, and lighting. It also features multi-reference support, enabling the integration of up to 10 reference images to maintain character consistency or stylistic themes without the need for specialized fine-tuning. The model is capable of precise adherence to complex instructions, including specific HEX color codes and pose requirements.
For optimal performance, users are advised to utilize a guidance scale between 2.0 and 3.5. Lower values (2.0-2.5) are recommended for achieving natural photorealism, while higher values (3.5+) are better suited for stylized or illustrated outputs. Because the model is distilled, it is most efficient when used at lower step counts, typically producing production-ready results in 4 to 6 steps.