The core idea, in the order it happens
Training runs forwards. Take a real image, add a little noise, add a little more, and keep going until nothing is left but static. At each level the network is shown the noisy image and asked to predict the noise that was added. That is the whole training objective, and it is much easier to optimise than an adversarial game, which is why this approach scaled to billions of images.
Generation runs the same ladder backwards:
- Start from a field of random noise, fixed by the seed.
- Ask the network what noise it sees, given the prompt conditioning.
- Subtract part of that prediction, landing on a slightly cleaner state.
- Repeat for the number of steps you set.
- Decode from latent space back to pixels.
Two consequences matter in daily use. First, the picture is decided early: composition and subject settle in the first few steps, and later steps mostly resolve texture. That is why a bad layout cannot be fixed by adding steps. Second, because there are many steps, there are many opportunities to intervene, which is where every control you have comes from.
What each setting is actually doing
| Setting | Effect | Typical range |
|---|---|---|
| Steps | Number of denoising increments | 20 to 40, or 1 to 8 on distilled models |
| Guidance (CFG) | Prompt pull per step | 4 to 8 on most image models |
| Seed | The starting noise field | Fix it for reproducibility |
| Denoising strength | How far your input image is pushed back into noise | 0.2 to 0.4 for variations, 0.6+ to reinvent |
| Scheduler | Step size distribution and noise reinjection | Model dependent |
The two most common mistakes come from treating steps and guidance as quality dials. Steps are a convergence budget, not a quality slider: once the image has converged, more steps produce an identical result at a higher cost. Guidance is a tradeoff, not a strength setting: pushing it up buys prompt adherence and pays in contrast, saturation and hard edges.
Video diffusion is a different problem
A video diffusion model has to denoise a stack of frames that agree with each other. Independent per-frame denoising would give you flicker, so these models add temporal layers, treat time as a third compression axis in the autoencoder, or use a transformer backbone that attends across frames as well as within them.
That is why video generation is so much more expensive, why clips are capped at a handful of seconds, and why the frames near the end of a clip drift more than the frames near the start. It also explains the sharpest control lever available: supplying a first frame removes the appearance decision from the model entirely and leaves it to solve motion only.
How to read all this into practice
Match the setting to the model, not to a rule of thumb. A turbo variant at 30 steps and CFG 8 will look worse than the same model at 6 steps and CFG 2. Model cards exist for a reason.
Diagnose before you re-roll. Soft or unresolved means too few steps. Burnt and oversaturated means guidance too high. Off-prompt means guidance too low or a prompt too long for the text encoder. Wrong composition means seed or a missing control input, not settings.
Change one thing at a time, with the seed fixed. With the seed left random, every comparison is confounded by a new noise field, which is the fastest way to draw a false conclusion about a setting.