cinemasetfreeThe parameters
Defaults are tuned per model and are usually right. The two worth learning are guidance scale (how literally the prompt is followed) and inference steps (quality against time and cost).
Image parameters
| Parameter | What it does | Too high looks like |
|---|---|---|
| Guidance scale | How strictly the image follows the prompt. Low is loose and often more natural; high is literal. | Over-saturated, over-contrasty, weirdly rigid compositions. Fried. |
| Inference steps | How many denoising passes. More steps means more detail, more time and more cost. | Diminishing returns; past a point you pay for changes nobody can see. |
| Scheduler | The sampling algorithm. Different schedulers converge differently; some need fewer steps. | Not a quality dial but a character dial. Try a few, keep what suits the look. |
| Aspect ratio | The output shape. Scenes generate 16:9; character portraits are portrait-shaped. | Asking for a ratio far from what the model was trained on distorts composition. |
| Prompt strength / Strength | In image-to-image, how far to move from the input. Low keeps the input, high ignores it. | At maximum the reference image stops mattering, which defeats the point of using one. |
| Clip skip | How many layers of the text encoder to skip. A stylistic lever inherited from Stable Diffusion. | Leave at the default unless a LoRA’s documentation asks for something else. |
| Scale factor | On upscale models, the multiplier applied to the source resolution. | Large factors invent detail that wasn’t there. 2× is usually honest; 8× is a guess. |
Video parameters
| Parameter | What it does |
|---|---|
| Duration | Clip length in seconds. Cost scales with it roughly linearly, and most models degrade past their trained length. |
| FPS | Frames per second of the generated clip. Higher is smoother and proportionally more expensive. |
Every generation costs credits. Test parameter changes on one image before running a bulk generate across the whole cast.