Plain-English explanation
Diffusion generation doesn't produce an image in one shot; it takes repeated passes, each one nudging a noisy latent image a little closer to a finished result. Inference steps is the number of passes run before the image is decoded and returned. A generation set to 20 steps runs 20 denoising iterations; one set to 50 runs 50.
How it works
At each step, the model predicts the noise in the current latent, and the sampler uses that prediction (following the noise schedule laid out by the scheduler) to compute a less-noisy version for the next step. This repeats for the configured step count until the schedule reaches its end. Because image-to-image's "strength" setting works by determining how many of the total steps are used (a lower strength skips the earlier, noisier steps and starts partway through), inference steps and strength interact directly in that mode: the same nominal step count applies differently depending on strength.
Adult-AI use
Step count is a slider present in essentially every adult-AI image generator, and it's one of the first settings people learn to tune since it trades directly against generation time and, on hosted platforms, sometimes against credit or compute cost. Finding the lowest step count that still produces clean results for a given model and sampler combination is a common optimization, especially for anyone generating many variations of a character in one session.
How it differs from sampler, scheduler, and guidance scale
Inference steps is purely a count; it says nothing about which algorithm runs at each step (that's the sampler) or how noise is distributed across those steps (that's the scheduler). It's also unrelated to guidance scale, which controls how strongly each step follows the prompt rather than how many steps run. All four settings are commonly adjusted together, but changing one doesn't change what the others do.
Limitations and misconceptions
- More steps does not automatically mean better output. Per the house convention on this technique: quality gains diminish past a model-dependent point, and pushing step count very high can plateau, waste compute, or in some cases subtly degrade the image.
- The "right" step count depends heavily on the sampler and model. Some releases use distillation to target a short schedule and should be run near their documented range rather than an older model's step count.
- Doubling the step count roughly doubles generation time (for samplers that don't share work across steps), so it's not a free quality dial.