피파 한 줄 정리: Steps는 더 많을수록 좋은 게 *아니야*. 30-50 근처에서 plateau가 와. Schnell/Flash 같은 모델은 4-8 steps로 충분하고.
When you generate an image, one of the key parameters is sampling steps (sometimes called "inference steps" or just "steps"). This controls how many denoising increments the model takes to go from pure noise to a clean image. It's a direct tradeoff between speed and quality — but the relationship isn't as simple as "more = better."
Think of it like sketching. If you give an artist 10 seconds, you get a rough gesture drawing — captures the essence but lacks detail. Give them 30 seconds, you get a solid sketch with proportions and shading. Give them 5 minutes, you get refined detail. But give them an hour for the same sketch? You get diminishing returns — past a point, more time just means obsessive noodling over already-good areas.
What Happens at Different Step Counts
Steps: 5 10 20 30 50 100
┃ ┃ ┃ ┃ ┃ ┃
Quality: 😰 🤔 😊 😄 😄 😄
Rough Okay Good Great Same Diminishing
shapes forms detail quality quality returns
Speed: ⚡⚡⚡⚡⚡ ⚡⚡⚡⚡ ⚡⚡⚡ ⚡⚡ ⚡ 🐌
Fastest Slowest
Here's the practical reality for most modern models:
- 5-10 steps: Quick previews. Useful for rapid iteration — "does this prompt direction work at all?" Composition is visible but details are rough.
- 15-25 steps: The sweet spot for most workflows. Good quality, reasonable speed. FLUX Schnell (the fast variant) is designed for 4-8 steps; FLUX Dev works great at 20-30.
- 30-50 steps: High-quality output. Marginal improvement over 20-25 for most prompts, but can help with complex scenes or fine details.
- 50+ steps: Rarely necessary. You're paying in generation time for quality improvements that are often invisible to the human eye.
Schedulers Change Everything
The "step schedule" — how much noise gets removed per step — is controlled by a scheduler (or sampler). Different schedulers distribute the work differently:
- Some front-load major changes in early steps (faster convergence)
- Some distribute work evenly (smoother refinement)
- Some are optimized for fewer total steps (like DPM++ 2M or Euler A)
The scheduler matters as much as the step count. A good scheduler at 20 steps often beats a poor scheduler at 50 steps.
- Sampling steps control the speed vs. quality tradeoff — but more isn't always better.
- The sweet spot for most models is 20-30 steps. Below 10 is for previews; above 50 rarely helps.
- The scheduler (sampler) matters as much as the step count — it determines how work is distributed.
- Use low steps for exploration, high steps for final outputs. Don't waste time rendering previews at maximum quality.