Look at the settings your creative team uses for AI image generation. If the step count is 80 or 100 and the guidance scale is in the teens, you are paying for a superstition: the belief that more computation always means more quality.
It does not, and the cost compounds across every render, every person and every project.
What the Research Says
A diffusion model builds an image over a series of steps, and each step is a full pass through the network. Two influential papers changed how many steps a good image needs:
- DPM-Solver++ reports high-quality guided samples "within only 15 to 20 steps", against the 100 to 250 that the older DDIM sampler typically needs.
- Karras and colleagues (NeurIPS 2022) reached state-of-the-art results on their benchmarks with 35 network evaluations per image.
The guidance setting has its own trade-off. Its inventors describe classifier-free guidance as trading sample fidelity against variety. Higher values follow the prompt more literally and give less diversity. They are not free quality.
The Cost in Plain Numbers
Render time scales almost directly with steps. The PyTorch team's SDXL baseline took 7.36 seconds for 30 steps on an A100. Apply that proportion to a team:
| Steps per image | Relative time | Hours per 10,000 images at 7.36 s per 30 steps |
|---|---|---|
| 25 | 1.0x | about 17 |
| 50 | 2.0x | about 34 |
| 100 | 4.0x | about 68 |
On slower workstations the absolute hours grow; the ratio does not. A team defaulting to 100 steps spends four times the machine time of one at 25, often for differences nobody can see.
Why the Habit Persists
If the research has been clear for years, why do teams still default to 80 or 100 steps? Three reasons, and none of them is technical.
- Settings travel without context. A workflow shared online with high values gets copied, and the copied values become a team's default without anyone testing them.
- The cost is invisible. Machine time does not appear on an invoice. Nobody signs off on an extra 30 seconds per image, so nobody questions it.
- Higher feels safer. When a deliverable matters, turning every dial up feels like diligence. It is the creative equivalent of over-provisioning a server just in case.
Each of these is a management problem, which is why the fix is a standard rather than a tip.
Size Multiplies the Waste
Steps are half of the cost. Every step processes the whole image, so larger outputs make each step more expensive too. Doubling an image's width and height quadruples its pixels, and with them the work in every step. A team that combines high step counts with large drafts is paying a multiplied penalty on work that will mostly be discarded.
The practical rule: explore at small sizes with tested default settings, then render finals at full size once the composition is approved. Most creative teams already work this way with traditional tools, sketching before painting. AI workflows deserve the same discipline.
Guidance Deserves the Same Scrutiny
The guidance setting is the second dial teams over-turn. Because higher guidance trades variety for literal prompt-following, a team that pushes it high everywhere often gets images that look alike and look over-processed. That leads to more re-rolls, which brings the time cost back. Test guidance alongside steps, and set the lowest value that still follows the brief.
Replace the Habit With a Standard
- Test once per model. Fix the seed and prompt, render at 15, 20, 25, 35 and 50 steps, and find the first result that matches 50.
- Set that as the team default, and document it like any other production standard.
- Make higher settings an explicit choice for final assets, not a reflex for drafts.
- Revisit when models change. A new model resets the test, not the habit.
Good Defaults Are a Product Decision
Tools shape habits. In Vision Studio, our open-source generation app, every model starts from 25 steps at a guidance of 7.5, steps are capped at 50, and an iteration tree records exactly which setting changed between versions. The goal is that a quality change is always a deliberate decision someone can see.
A Worked Example
Take the published baseline of 7.36 seconds for 30 steps, about 0.25 seconds per step. A team producing 1,000 images a day at 100 steps spends about 6.8 hours of machine time daily. The same team at 25 steps spends about 1.7 hours. That is roughly five hours of compute a day, every working day, returned for a setting change that, after the test above, nobody can see in the output.
On slower workstation hardware the absolute hours are larger, so the saving is too.
The same logic applies to any AI setting that trades time for theoretical quality: upscaling passes run on every draft, batch sizes larger than anyone reviews, or video clips rendered longer than the edit will use. Each is a small habit with a compounding cost, and each yields to the same method: measure where the returns flatten, set the default there, and make anything above it a deliberate exception.
How to Roll It Out
Standards stick when they are easy to follow and easy to check. Publish the tested defaults per model in the team's playbook. Build them into templates so the default is the right value, not a value someone has to remember. Review a sample of final assets each quarter, noting the settings used, and retest when a new model enters the toolkit. The aim is not to forbid high settings. It is to make them a choice with a reason.
The Leadership Lesson
AI tooling invites "more is better" thinking because the costs are invisible: machine time, not invoices. Make them visible, measure where the returns flatten, and standardise there. The recovered hours go back into the work that actually differentiates.
The hands-on sampler and settings guide is on the Vision Studio blog at vision-studio-x.com/blog/more-steps-is-not-more-quality. Vision Studio is a free, open-source, local-first AI generation app built by Strategia-X. If you want a second pair of eyes on how your organisation adopts AI tooling, talk to us at strategia-x.com.
-Rocky
#CreativeOperations #Productivity #GenerativeAI #ProcessImprovement #AIStrategy #OpenSource #StrategiaX #VisionStudio