Inference Steps

Also known as: num_inference_steps, inference_steps, steps, sample_steps

Inference steps control how many denoising iterations the model performs. More steps generally mean higher quality but slower generation. Most models hit diminishing returns between 25 and 50 steps.

What It Does

Diffusion models generate images by starting from pure random noise and iteratively refining it into a coherent image. Each inference step removes a small amount of noise, guided by your text prompt. The process is like sculpting: the first few steps carve out the broad shapes and composition, while later steps refine textures, edges, and fine details.

With too few steps, the image looks blurry, noisy, or lacks fine detail because the model hasn’t had enough iterations to fully resolve the image. With enough steps, the image converges to its full quality and additional steps beyond that point add generation time with little to no visual improvement. The exact convergence point depends on the model architecture and the scheduler algorithm being used.

Some modern models are specifically designed to produce good results in very few steps. Flux Schnell, for example, uses a distilled architecture that generates quality images in just 1 to 4 steps, while standard models like Flux Dev or SDXL typically need 20 to 30 steps. Knowing your model’s sweet spot saves both time and credits.

Adjusting It in CSF

Step count is configured per production design. Edit a design, open its Image tab (or Video tab for video models), and click Advanced Settings to open the Advanced Model Settings modal. Choose the model to configure; CSF loads that model’s parameters, and inference steps appears only if the model exposes it — under whichever name it uses (num_inference_steps, steps, and so on).

It usually renders as a whole-number slider that runs between the model’s minimum and maximum, with the current step count shown beside it. Save stores your value for this design and model (only if it differs from the default), and Reset to Defaults restores the model’s recommended step count.

Value Ranges

Low (1 – 10)

Very fast, rough results. Useful for quick previews or specialized fast models like Flux Schnell, which is optimized for 4 steps. Standard models will look noticeably unfinished at this range.

Mid (20 – 30)

The sweet spot for most models. Good quality with reasonable generation speed. This is where the majority of detail emerges without excessive wait times.

High (40 – 100)

Marginal quality gains at the cost of significantly longer generation times. Can be useful for fine detail work or specific schedulers that benefit from more steps, but generally not worth the extra wait.

Tips

25 steps is a safe default for most diffusion models.
Some models like Flux Schnell are designed for as few as 1 to 4 steps.
If you’re iterating on prompts, use fewer steps for quick drafts, then increase for the final image.
Doubling steps rarely doubles quality — there are diminishing returns after 30 steps on most models.
Step count and scheduler are linked: a better scheduler can reach the same quality in fewer steps.