Strength

Also known as: strength, denoise

Strength determines how much an input image is transformed during generation. At 0, the original image is untouched. At 1.0, the model generates an entirely new image, ignoring the original. This parameter is key for image-to-image workflows.

What It Does

When a model works from a reference image alongside your prompt (image-to-image generation), it partially noises the input image, then denoises it according to the text prompt. The strength parameter controls how much noise is added. A low strength preserves most of the original composition, colors, and structure, making only subtle adjustments. A high strength gives the model more creative freedom, potentially departing significantly from the source image.

This matters for reference-driven workflows like style transfer and iterative refinement, where you want to nudge an existing image toward a new look without discarding what already works.

Adjusting It in CSF

Strength is a model-level parameter, so you set it in the Advanced Model Settings modal. Edit a production design, open its Image tab, and click Advanced Settings; then pick the model to configure. CSF loads that model’s parameters, and strength appears only if the model exposes it — many text-to-image models don’t, since it only applies when there’s an input image to transform.

When present it renders as a slider from 0 to 1 with the current value shown beside it. Save keeps your value for this design and model (only when it differs from the default), and Reset to Defaults restores the model’s default.

Value Ranges

Low (0.1–0.3)

Subtle adjustments. Original composition and colors are largely preserved. Good for minor touch-ups or gentle style changes.

Mid (0.4–0.6)

Noticeable transformation while maintaining the overall structure. Ideal for style transfers or moderate creative adjustments.

High (0.7–1.0)

Heavy transformation. The model may discard most of the input image’s structure. Useful when you want a fresh take inspired loosely by the original.

Tips

Start at 0.5 for image-to-image and adjust up or down based on results.
If results look muddy or half-baked, you likely need a higher strength.
Combining low strength with more inference steps can help preserve detail while still following the prompt.
Some models expose this as denoise instead — it’s the same control.