FLUX.1 [Dev]

The standard variant generates imagery using the base model's native photorealistic aesthetic, whereas the LoRA variant integrates custom style adapters to enforce specific artistic aesthetics like watercolor, vector art, or branded character design. LoRA variants run at a 48.3-credit rate compared to 34.5 credits for standard generation due to initializing custom adapter weights.

Use image-to-image when you need to preserve the compositional layout, silhouette, or geometry of an existing source image while altering styling, subjects, or textures. Standard text-to-image generation builds compositions entirely from descriptive prompts, offering unconstrained creative freedom from scratch.

A guidance scale between 3.5 and 4.0 delivers balanced prompt adherence, while lowering the scale to 1.5–3.0 enhances creative interpretation and photorealism. Increasing guidance to 5.0–10.0 enforces strict alignment for typography and vector graphics, though exceeding 7.0 risks over-saturation and visual artifacts.

The default setting of 28 inference steps provides the ideal balance between generation speed, skin texture clarity, and detail fidelity. For complex scenes featuring multi-layered reflections, intricate glass refractions, or tiny legible text, increasing the step count up to 50 yields maximum clarity.

Descriptive natural prose that outlines subject hierarchy, materials, camera specs, and volumetric lighting yields the highest quality. Avoid comma-separated keyword tag soup, negative-prompt phrases, or weighted parentheses, as the flow transformer interprets all syntax literally.