Open any shared workflow and look at the settings. Steps: 80. CFG: 14. Sometimes 150 steps, because if 50 is good then three times as many must be three times as good. It is the most common superstition in image generation, and it costs real time on every single render.
Steps are not a quality dial. They are a time dial with a quality curve attached, and that curve goes flat much earlier than most people think.
What a Step Actually Is
A diffusion model starts from noise and removes it over a sequence of steps. Each step is a full pass through the denoising network. The sampler, also called the scheduler, decides how big each move is and how the next one is computed. Twice the steps means roughly twice the network passes, which means roughly twice the time.
The question is never "are more steps better?" in the abstract. It is "has this sampler already converged?", and for modern samplers the answer arrives early.
What the Research Measured
Two papers reshaped how many steps a good image needs.
In DPM-Solver++: Fast Solver for Guided Sampling of Diffusion Probabilistic Models, Cheng Lu and colleagues report high-quality guided samples "within only 15 to 20 steps", against the 100 to 250 steps the older DDIM sampler typically needs. DPM-Solver++ is the "DPM++" in the sampler names you see in most apps.
In Elucidating the Design Space of Diffusion-Based Generative Models (NeurIPS 2022), Tero Karras and colleagues reached state-of-the-art image quality on their benchmarks with 35 network evaluations per image, and contributed the noise schedule that gives "Karras" sampler variants their name.
Neither paper says your particular model and prompt will look identical at 20 steps and at 80. Both say the bulk of the work is done far earlier than the habits suggest.
Same Names, Same Math
Sampler names travel between apps, which helps. The Diffusers scheduler documentation maps the familiar names onto its implementations: "DPM++ 2M" is the multistep DPM-Solver scheduler, and "DPM++ 2M Karras" is the same scheduler initialised with Karras sigmas. If you learned a sampler in one tool, the knowledge carries.
CFG Is a Trade, Not a Volume Knob
Classifier-free guidance, the CFG scale, sets how hard the model is pushed toward your prompt. Jonathan Ho and Tim Salimans, who introduced it in Classifier-Free Diffusion Guidance, describe it as a trade between sample fidelity and mode coverage, in the same spirit as low-temperature sampling.
Translated: higher guidance follows the prompt more literally and gives you less variety. Push it far enough and images tend to look overcooked: harsh contrast, saturated colour, brittle detail. Pushing CFG is not "more prompt adherence for free". You pay for it in diversity, and often in looks.
The Cost Side of the Ledger
Time is the part you can measure precisely. The PyTorch team's SDXL benchmark clocked 7.36 seconds for a 30-step image on an unoptimised A100 baseline. Because every step is another pass through the network, steps scale render time almost directly:
| Steps | Relative render time | What it usually buys with a modern sampler |
|---|---|---|
| 15 to 20 | 0.6x to 0.8x | Converged composition, enough for drafts and most finals |
| 25 | 1.0x | A safe default across model families |
| 50 | 2.0x | Occasional fine-detail gains on some models |
| 100 | 4.0x | Rarely anything visible over 50 |
The middle column is arithmetic. The right-hand column is a rule of thumb to test against, not a law, which is the point of the next section.
Size Changes the Price of a Step
Steps are only half the cost. Every step processes the whole latent image, so a larger render makes each step more expensive as well. In Vision Studio the resolution tiers set the long edge to 512, 768 or 1024 pixels, and custom sizes run from 256 to 2048 pixels a side. Going from the 512 tier to the 1024 tier at the same aspect ratio quadruples the pixel count of every step.
That interaction is why the habit of high steps hurts most at high resolution. Eighty steps on a small draft is wasteful. Eighty steps on a large final is wasteful at four times the price. Find the knee at draft size, where experiments are cheap, then confirm it once at final size rather than assuming the large render needs more.
Find Your Own Knee
The convergence point varies by model, sampler and subject. You can find it in a few minutes:
- Fix the seed. Same prompt, same seed, same size. Now only steps vary.
- Render at 15, 20, 25, 35 and 50 steps. Put them side by side.
- Find the first one you cannot tell from 50. That is your knee for this model and sampler.
- Repeat for CFG. Hold steps at your knee and try 4, 6, 7.5, 9 and 12. Pick the lowest that still follows the prompt.
Ten minutes of this saves hours across a project, and it replaces a superstition with a measurement on your own hardware.
How Vision Studio Sets It Up
Vision Studio exposes eight samplers (Euler, Euler a, DPM++ 2M, DPM++ 2M Karras, DPM++ SDE, DPM++ SDE Karras, DDIM and UniPC), steps from 1 to 50, CFG from 1 to 20, and a fixed or random seed.
Every model starts from the same defaults, 25 steps at CFG 7.5. That is deliberate. We do not tune hidden per-model values behind your back, so a change you see in the image is a change you made. The 50-step ceiling is deliberate too. If a model needs more than 50 steps to converge with a modern sampler, more steps are not the fix.
The iteration tree makes the sweep above painless: fork a render, change one value, and each child records its settings difference against its parent, so you can see exactly what moved.
Spend the Seconds Elsewhere
Every unnecessary step is time you could spend on a better prompt, a ControlNet guide or one more variation. Find the knee for the models you use, set your defaults there, and treat anything above it as a deliberate choice, not a reflex.
Vision Studio is free, open-source (MIT), and runs locally on your GPU. The sampler list, ranges and defaults are documented with their source files at vision-studio-x.com/features. Download it at vision-studio-x.com/download.
-Rocky
#Diffusion #Samplers #StableDiffusion #AIArt #Workflow #OpenSource #VisionStudio