Three settings appear in almost every AI video tool that offers any control: the seed, the guidance scale (often called CFG) and the number of steps. Understanding them turns generation from a lottery into a controllable process. This guide explains each setting, what changing it does and how to find good values for your model.
Seed: controlling randomness
Every generation starts from random noise. The seed is the number that determines that noise. Same seed, same prompt, same settings, same model: same result. Change the seed and you get a different composition, even with an identical prompt.
How to use seeds
- Exploration: use random seeds to see a variety of interpretations of your prompt.
- Refinement: once you find a promising result, fix its seed. Now changes you make to the prompt or settings show their real effect.
- Reproduction: save the seed with your keepers so you can regenerate at higher quality later.
Note that seeds are only reproducible on the same model, settings and often the same software and hardware type. Different GPUs or software versions can produce slightly different results from the same seed.
Guidance scale (CFG): how closely to follow the prompt
Classifier-free guidance controls how strongly the model is pushed towards your prompt and away from your negative prompt at each step.
- Low guidance: the model follows its own sense of what looks natural. Results are softer and more varied, but may ignore parts of your prompt.
- Medium guidance: the usual sweet spot: prompt adherence with natural images.
- High guidance: strong prompt adherence, but images can become over-contrasted, over-saturated or “burnt”, and motion can become stiff.
Good values differ a lot between models. Many video models work best in a moderate range, often somewhere between about 3 and 8, while some distilled or fast models are designed to run at a guidance of 1 or with guidance disabled. Always start with the value recommended in the model’s documentation or example workflow.
Steps: how many refinement passes
Steps are the number of denoising iterations. Each step makes the video a little clearer.
- Too few steps: blurry, noisy or unfinished-looking results.
- Enough steps: clean results. For many models, 20 to 50 steps is typical.
- More steps: diminishing returns. Beyond a point, results barely change while generation time keeps growing.
Some fast or distilled models are designed for very few steps, sometimes 4 to 8. Running them with 30 steps does not help and can even hurt. Again, follow the model’s recommendation.
How they interact
Steps and guidance work together. Higher guidance can need slightly more steps to resolve cleanly. The sampler (the mathematical method used for denoising) also affects how many steps are needed. When you change the sampler, re-test your step count.
Other settings worth knowing
- Sampler and scheduler: different methods for removing noise. Use the model’s recommended one first; experiment later.
- Shift or flow shift: in some recent models, a setting that affects how detail and motion are balanced across steps. Follow example values.
- Denoise strength: in image-to-video and video-to-video, how much the input can change.
A testing method
To learn a new model quickly, run a small grid test:
- Choose one prompt and one seed.
- Generate with the recommended guidance and steps. This is your baseline.
- Generate with guidance lower and higher than the baseline (for example −2 and +2), keeping steps the same.
- Generate with steps at half and one-and-a-half times the baseline, keeping guidance the same.
- Compare all clips side by side.
Five generations teach you how your model responds. Repeat with a different type of prompt, such as a portrait instead of a landscape, because responses can vary by subject.
Recommended habits
- Draft with fewer steps and lower resolution; render finals with the best settings.
- Always note seed, guidance, steps and sampler for keepers.
- Change one setting at a time.
- Do not assume that more is better for steps or guidance.
- Re-test when you update the model or software.
Common problems and the setting to check
| Problem | Check |
|---|---|
| Prompt details ignored | Guidance slightly higher |
| Harsh, oversaturated look | Guidance lower |
| Blurry, unfinished result | More steps, correct sampler |
| Generation too slow | Fewer steps, lower resolution for drafts |
| Cannot reproduce a result | Seed, model version, all settings saved? |
| Stiff motion | Guidance lower |
Batch seeds for choice
When a prompt is close to right, a good habit is to queue a batch of six to ten seeds at draft quality and pick the best. Randomness is not your enemy; it is a source of variety. A single excellent seed is often worth more than an hour of tweaking guidance by small amounts. Once you pick a winner, lock the seed and do your fine-tuning from there.
Settings on hosted tools
Many hosted generators hide these settings or rename them, for example “creativity”, “prompt strength” or “quality”. “Creativity” usually means lower guidance; “quality” often means more steps or a larger model. If a tool shows a seed, use it as described here; if not, save your best clips immediately, because you may not be able to recreate them.
A note on reproducibility for clients
If you produce work for clients, reproducibility is a professional requirement. A client may approve a draft and ask for the final in higher resolution weeks later, or request a small change to one shot. Store the seed, all settings, the model version and the workflow file for every approved draft. Without them, you may find yourself unable to recreate the exact shot the client fell in love with, which is an awkward conversation to have.
Frequently asked questions
Is there a best seed?
No. Seeds are random starting points; a good seed for one prompt means nothing for another.
Why does the same seed look different on another computer?
Different hardware, software versions or precision settings can change results slightly.
Should I always use the maximum steps?
No. Use the lowest value that gives clean results, and save time for more variations.
