Generation controls

    Sampling steps

    Also called Inference steps, Denoising steps.

    Sampling steps is how many passes a model takes to turn its starting noise into a finished output — more passes, more refinement, more time.

    Generation is iterative. Each step removes a little of the noise and re-estimates what is underneath, so the picture resolves gradually rather than appearing at once. The step count is how many of those refinements you are paying for, and it maps almost linearly onto how long the job takes.

    Returns flatten fast. The first handful of steps do the structural work — where things are, how big they are — and later steps mostly polish texture. Past the model's comfortable range the output stops improving and can start drifting, so a very high number is usually wasted time rather than insurance.

    The right number depends on the sampling method and the model. Distilled variants are specifically trained to converge in very few steps, which is the entire trick behind fast models: fewer passes, not smaller output.

    In practice

    • Structure is decided early; detail is decided late. If the composition is wrong, more steps will not fix it.
    • Time scales roughly with the step count, so it is the first thing to cut when drafting.
    • Fast and turbo variants are tuned for a low step count and get worse if you raise it.

    The mistake to avoid

    Treating steps as a quality slider and pushing it to the maximum. Most models plateau well before that, and you have bought minutes rather than pixels.

    Related terms

    The all-in-one AI studio for creators. 60+ models for video, image, voice, music and lipsync in a single app.