The fast local model: fifteen seconds a frame, and what it gives up

At three minutes a frame, every generation is a commitment. You write the prompt carefully, you wait, you evaluate. Iteration is expensive enough that you plan around it.

At fifteen seconds, that changes qualitatively. You can try a phrasing, see it, adjust, see it again — three times in a minute. Generation stops being a production step and becomes a way of working out what you want.

That difference is worth more than most quality comparisons suggest, and it is why a fast model earns a place in the setup even when a slower one produces better final frames.

Fast models produce competent images: composition is sound, subjects are coherent, prompts are followed. What tends to be missing is the surface irregularity that makes a photograph read as photographed — grain structure, subtle falloff, the imperfection of a real lens.

The result is a digital cleanliness. For a great deal of work that is fine or even preferable. For a frame that has to sit next to real photography without announcing itself, it is the thing that gives it away.

Judge it on your own material rather than on the general claim. The gap is smaller on some subjects than others, and on graphic or product work it can be negligible.

These models achieve their speed through a distilled sampling process that reaches a good result in a handful of steps. That process has an optimum, and it is low.

Raising the step count does not improve the image. It slows generation and frequently degrades the result — over-sharpened, with a hardness in the highlights.

If a frame is not working, the variable is the prompt or the seed. This is the single most common misconfiguration, carried over from habits formed on earlier models where more steps genuinely meant more quality.

Iterate on the prompt with the fast model until the composition, content and framing are right. Then run the locked prompt through the slower, more photographic model to produce the frame you will actually use.

This works because the two respond to prompts similarly enough that the exploration transfers. It will not be identical, and it does not need to be — you resolved the structural questions cheaply and spent the expensive generation on the answer.

Where the fast model's output is good enough on its own, use it. The point is having the choice rather than defaulting to the slow path for everything.

• Boards and drafts. Nobody is examining the surface of a board frame; they are looking at the idea.
• Volume. Hundreds of assets where individual character does not matter.
• Variation hunting. Producing forty options to find the composition worth developing.
• Anything being composited heavily. The photographic surface you paid three minutes for gets destroyed in post anyway.

Recognising these saves a great deal of time that would otherwise be spent generating carefully for a purpose that did not need it.

Keep both loaded in your workflow, with graphs ready for each, so choosing is a click rather than a reconfiguration.

Learn each one's prompt preferences separately. They respond differently to the same phrasing, and assuming a prompt transfers unchanged is a common source of the conclusion that one is worse.

The setup that works is not the best model. It is two models whose strengths you know, reached for without thinking about it.