FLUX.1 [Dev]

The standard variant generates pristine images directly from text prompts at 34.5 credits, while the LoRA variant enables injecting custom trained style or subject adapters at 48.3 credits. Choose the standard model for general photorealism and typographic scenes, and reach for the LoRA variant when you need to match specific brand design guidelines, unique illustrative genres, or custom character likenesses.

Use image-to-image when you have an existing visual composition, lighting setup, or product mockup that you want to transform or restyle with a prompt. By setting the strength parameter between 0.30 and 0.85, you can control whether the model performs subtle material touch-ups or executes a dramatic artistic overhaul while retaining structural layout.

Write descriptive, natural English sentences between 40 and 80 words detailing lighting, camera lenses, materials, and exact text inside quotation marks. Avoid keyword-stuffed strings and negative prompts, as the model's T5-based text encoder relies on narrative scene descriptions rather than comma-separated tag lists.

Set the guidance scale to 3.5 and inference steps to 28 for optimal balance between prompt adherence and visual fidelity. Setting guidance scale above 4.0 rapidly introduces oversaturated, high-contrast, or burned color artifacts, while steps can be increased to 40–50 when resolving ultra-dense macro textures or intricate typography.

FLUX.1 [Dev] was developed by Black Forest Labs and released on August 1, 2024. It is an open-weight, 12-billion parameter rectified flow transformer guidance-distilled from the closed-source FLUX.1 [Pro] for personal and commercial applications.