Generation Parameters
Each parameter has its own "personality". Understanding them helps you avoid pitfalls and save credits.
Not a pro? No problem
If you're not familiar with these parameters, you don't need to actively tweak them — the defaults are already a strong combination, and the output will be good enough out of the box.
If you accidentally break things, head to Settings and tap "Reset to default" to restore the system defaults in one tap.
Width × Height
Determines aspect ratio and resolution.
| Use case | Recommended size | Notes |
|---|---|---|
| Portrait (character art, person) | 832 × 1216 | Best for single subject, natural composition |
| Landscape (scenery, scene) | 1216 × 832 | Best for environment, group shots |
| Square (avatar) | 1024 × 1024 | Universal size, ideal for social avatars |
WARNING
Larger canvases cost more, and total pixel count drives cost more than dimensions per se. Unless you specifically need high resolution, the sizes above are recommended.
Count
How many candidate images to generate at once.
- Want to save credits: pick 1
- Want options to choose from: pick 2–4
Cost scales linearly with count — generating 4 costs roughly 4× as much.
TIP
When you're not sure whether your prompt is well-tuned, generate 1 first to test. Once stable, use 4 to get more compositions.
Steps
Number of sampling iterations the model runs. More steps mean finer details but higher cost.
| Steps range | Best for | Output traits |
|---|---|---|
| 20–24 | Quick previews, composition tests | Overall shape OK, details a bit rough |
| 28–32 | Daily use (recommended) | Sufficient detail, best value |
| 40+ | Extreme detail | Diminishing returns, often imperceptible |
Prompt Guidance
The "guidance strength" of your prompt. Higher values make the AI follow your prompt more strictly; lower values let it improvise.
| Range | Behavior |
|---|---|
| Below 3.0 | Too much improvisation — easily off-topic |
| 5.0 – 6.0 | Natural, picture-like (default recommended) |
| 7.0 – 8.0 | Strict to the prompt but tends to feel stiff |
| 9.0+ | Overexposure, color saturation issues, increased risk of breakdown |
Prompt Guidance Rescale
When Prompt Guidance is high (≥7.0), the output may suffer from overexposure or color distortion. Prompt Guidance Rescale is used to "soften" the over-strong guidance and produce more natural images.
- 0.0 (default): no softening
- 0.1 – 0.3: gentle softening, pairs with high Prompt Guidance
- 0.5+: heavy softening, abstract feel
If you keep Prompt Guidance in the 5.0–6.0 sweet spot, leave Prompt Guidance Rescale at 0.
Sampler / Noise Schedule
Controls the algorithmic path during generation. The defaults handle the vast majority of cases; advanced users can experiment with different combinations for varied feel.
Different Samplers have different preferences:
- Toward photoreal / stable → default Sampler + default Noise Schedule
- Toward anime / stylized → try ancestral Samplers (those marked
_aorancestral) - For sharper detail → karras-family Noise Schedules typically perform better
TIP
If you're not familiar, don't change these two — many combinations are mismatched, and random tweaks may break the output.
Seed
The "password" that controls randomness.
- Empty / random: every run produces a different result (recommended)
- Specific number: same prompt + same Seed ≈ consistent output (reproducible)
Common technique
Lock the Seed + tweak the prompt: after finding an image you like, fix its Seed, then change just one keyword in your prompt (hair color, outfit, etc.) to get "the same character, different versions" series.
VarietyPlus
Increases diversity across multiple images in the same batch (composition, pose, viewpoint).
- Off: similar outputs in the batch — easier to pick the most stable version
- On: more variety in the batch — better for exploring compositions
When generating only 1 image, this toggle has no effect. For 2–4 images, turning it on prevents the "all 4 look identical" problem.
Tips
- First time: keep all parameters default; tune the prompt first
- Once stable: tweak canvas size and count
- Don't change multiple at once: change one at a time to identify what's making the difference