Seeds, Samplers, and CFG: The Settings That Actually Matter
Why Settings Matter
You've got a checkpoint. You've got a prompt. You click generate and something comes out. But every time you click, you get a different image — sometimes great, sometimes terrible.
That's not random. It's controlled by a handful of settings. Most of them you can ignore, but a few of them make a real difference. Here's what they do.
Seed: Your Image's DNA
The seed is a number that determines the starting point for generation. Same prompt + same seed + same settings = the exact same image every time.
When the seed is set to -1 (the default), it picks a random number each time. That's why you get a different image every click.
Why you'd want to lock a seed:
- You generated something you love and want to make small tweaks without losing it. Lock the seed, change one thing in the prompt, and the image stays mostly the same.
- You want to compare how a specific composition looks across different checkpoints or settings. Same seed keeps the composition consistent.
How to use it: After you generate an image, the seed is shown in the output info. Copy it and paste it into the seed field. Now that image is reproducible.
Important: The seed only gives you the same image if EVERYTHING else is identical — same checkpoint, same prompt, same size, same steps, same sampler, same CFG. Change any of those and the seed produces something different. This is actually why copying someone's prompt from Civitai doesn't give you their image even if you copy their seed — your checkpoint or settings are probably different.
Steps: Quality vs Speed
Steps is how many times the AI refines the image. More steps = more detail, up to a point.
Here's the reality:
| Steps | What You Get |
|---|---|
| 1-10 | Blurry mess. Useless. |
| 15 | Rough but recognizable. Good for quick tests. |
| 20 | Good enough for most images. This is my default for scouting. |
| 25-30 | Sweet spot. Noticeable quality bump over 20. |
| 40 | Marginal improvement. Slower. |
| 50+ | Wasting time. You won't see the difference. |
My approach: I generate scouting batches at 20 steps to find compositions I like. When I find winners, I regenerate them at 30 steps for the final version. Going higher than 30 is almost never worth it.
Sampler: The Generation Algorithm
The sampler is the math behind how the AI builds the image step by step. There are a lot of them. Most people overthink this.
The ones that matter:
- Euler a — fast, creative, slightly unpredictable. Good for exploration. This is what I use most of the time.
- DPM++ 2M Karras — clean, detailed, consistent. Good when you want reliable results.
- DPM++ SDE Karras — similar to 2M but with more variation between steps. Good for artistic styles.
The ones you can ignore for now: DDIM, UniPC, LMS, Heun, DPM adaptive, and the 20 others. They exist for specific use cases but you won't need them starting out.
My recommendation: Start with Euler a. If your images look too chaotic or inconsistent, switch to DPM++ 2M Karras for more control. That covers 90% of use cases.
CFG Scale: How Literally to Follow Your Prompt
CFG (Classifier-Free Guidance) scale controls how closely the AI follows your prompt. It's a number, usually between 1 and 30.
- Low CFG (1-4): The AI interprets your prompt loosely. More creative, more unexpected, sometimes more natural-looking. But it might ignore parts of your prompt.
- Medium CFG (5-8): The sweet spot. Follows your prompt while still having room to make it look good.
- High CFG (9-15): Follows your prompt very literally. Colors get more saturated. Contrast increases. Can look overdone or artificial.
- Very high CFG (15+): Usually looks terrible. Colors blow out, artifacts appear, everything looks deep-fried.
My default: 3-5 depending on the checkpoint. Newer SDXL checkpoints tend to work best at lower CFG (3-4). Older SD 1.5 checkpoints often need 7-8.
How to find the right CFG for your checkpoint: Generate the same prompt at CFG 3, 5, 7, and 10. Compare. You'll quickly see where the sweet spot is. Once you find it, leave it there.
Resolution: Size Changes Everything
This isn't just "how big is my image." Resolution changes the composition.
- Tall (512x768 or 832x1216) — portraits, single characters, vertical compositions
- Wide (768x512 or 1216x832) — landscapes, scenes, horizontal compositions
- Square (512x512 or 1024x1024) — centered subjects, symmetrical compositions
SD 1.5 checkpoints are trained on 512x512. Use multiples of that: 512x768, 768x512, etc. Going much larger than 768 on either side without upscaling usually causes problems (duplicate heads, repeated patterns).
SDXL checkpoints are trained on 1024x1024. Use 832x1216, 1216x832, 1024x1024, etc.
Don't generate huge images directly. If you want a 2048x3072 wallpaper, generate at 832x1216 first, then upscale. Generating at full resolution often produces artifacts and takes forever.
The Settings I Actually Change
Here's my honest workflow:
| Setting | What I Do |
|---|---|
| Seed | Leave at -1 for exploration. Lock it when I find something I like and want to tweak. |
| Steps | 20 for scouting batches. 30 for final versions. Never higher. |
| Sampler | Euler a for most things. DPM++ 2M Karras when I want consistency. |
| CFG | 3 for my current SDXL checkpoints. I found the number and I leave it alone. |
| Size | 832x1216 portrait for most of my work. I change it when the composition calls for it. |
That's it. Five settings. Everything else stays at defaults. I don't touch the scheduler, I don't mess with eta, I don't change the VAE unless a checkpoint specifically needs one.
The biggest mistake beginners make with settings is changing too many things at once. Change one setting, see what it does, then change another. If you're tweaking CFG and sampler and steps all at the same time, you'll never know which change made the difference.
What's Next?
- Checkpoints vs LoRAs vs Embeddings — the three types of models
- How I Remix Any Prompt — turning someone else's prompt into something new
- The Three Starting Points — how I decide where to begin each image