Parameters
What are inference steps?
Inference steps are the number of denoising passes the model runs, trading generation time against how far the image is refined.
Each inference step is one pass where the model looks at the current noisy image, predicts the noise, and removes some of it. More steps means the path from static to image is walked in smaller increments, which generally means a more refined result. It also means proportionally more computation, so 50 steps takes roughly twice as long as 25.
The relationship between steps and quality is not linear, and this is the part most people get wrong. Improvement is steep at the low end and flattens quickly. Going from 10 steps to 25 usually makes an obvious difference; going from 50 to 100 usually makes almost none, while doubling the wait. Past a certain point the image has converged and extra steps are just re-confirming it.
Where that point sits depends on the scheduler. Efficient modern schedulers like DPMSolverMultistep reach a converged image in fewer steps than simpler ones, which is most of their advantage. With a good scheduler, 20 to 30 steps is often indistinguishable from 50 for a typical prompt.
The practical workflow is to iterate at a low step count and finalise at a higher one. Generate at 20 steps while you are testing wording and compositions, then re-run your chosen prompt and seed at 40 or 50 for the version you actually keep. Dreamify defaults to 50 steps, which is comfortably in converged territory for all three of its schedulers.
In practice
- 10 steps
- Fast, often visibly unfinished or soft
- 20 to 30 steps
- The sweet spot for iteration on a good scheduler
- 40 to 50 steps
- Converged, sensible for a final image
- 100+ steps
- Rarely distinguishable, twice the wait
Worth remembering
- Steps are denoising passes; more steps means more time
- Quality gains flatten out quickly, usually well before 50
- Better schedulers converge in fewer steps
- Iterate low, finalise high
Try it in the generator
The advanced panel exposes the negative prompt, scheduler, inference steps and seed. Twenty images free, no account needed.
Start generatingFrequently asked questions
- How many steps should I use?
- Around 20 to 30 while iterating, 40 to 50 for a final. Going much beyond 50 mostly costs time without changing the picture.
- Do more steps fix a bad prompt?
- No. Steps refine what the model is already heading toward. If the composition is wrong, the fix is the prompt or the seed, not more steps.
- Do steps cost more credits?
- On Dreamify each generation counts as one image regardless of step count. Higher step counts simply take longer to come back.
Related terms
What is a scheduler (sampler)?
The scheduler, or sampler, is the algorithm that decides how large each denoising step is and how the model moves from noise toward the final image.
What is a seed in AI image generation?
The seed is the number that determines the random starting noise, so the same seed and prompt reproduce the same image every time.
What is a diffusion model?
A diffusion model generates data by learning to reverse a noising process, turning random static into an image one denoising step at a time.
Back to the glossary, or apply it on a use case or an art style.