The whole clip is made at once, not frame by frame

Starting from noise, all at once

Generation begins with the entire latent block filled with random noise: all 30 latent frames of it. The model then makes a prediction about that noise, subtracts part of it, and repeats. After a fixed number of steps the block is clean and gets decoded.

The important word is simultaneously. At step 7, frame 1 and frame 120 are both 7 steps along. They are being refined together, seeing each other, agreeing with each other about where the wall is.

This is the opposite of how most people imagine it, and it explains several things that otherwise look arbitrary:

• You cannot watch the first second while the rest renders. There is no first second yet. Progress bars that show a preview are decoding a partially denoised block, which is why early previews look like coloured fog.
• You cannot generate a bit more onto the end. The block size was fixed before step 1. Extending means a new run conditioned on a decoded frame, with all the photocopy damage that implies.
• Cost is decided at the start. Duration, resolution and step count are all fixed before any picture exists. There is no way to spend more only on the interesting part.

What a step actually buys

Each step is one full forward pass of the transformer over the whole latent. Twenty steps means twenty passes; fifty means fifty. Cost is close to linear in steps.

Quality is not. The curve is steep at the start and flat after a point that is lower than most people expect. On a typical modern video model:

• 4 to 8 steps: shapes and colour are right, motion is mushy, fine structure is unresolved.
• 15 to 25 steps: most of the quality. This is where the majority of production work sits.
• 30 to 50 steps: small improvements in fine detail and edge stability.
• Beyond 50: you are usually paying for a different random walk, not a better one.

Run this yourself with a fixed seed. Generate the same shot at 10, 20, 30 and 50 steps and put the four side by side. Most people are surprised at how early the returns stop, and it is worth knowing your own model’s number rather than a general one.

Flow matching, and why older advice does not transfer

Early diffusion models used a noise schedule with hundreds of possible steps and samplers with names like DDIM and Euler-a. Most current video models are trained with flow matching instead, which learns a straighter path from noise to data and converges in far fewer steps.

This matters practically because a lot of advice written for 2022-era image models is wrong here. Step counts of 100, high guidance values, and sampler names that no longer exist in the interface all come from that lineage. If a tutorial tells you to use 80 steps on a flow-matching video model, it was written for different machinery.

Flow-matching models also expose a shift or sigma shift parameter, which redistributes where the steps are spent along the path. Higher shift spends more effort in the early, structural part of the denoise. Video models generally want a higher shift than image models, and the tool usually sets a sensible default. Change it last, if at all.

Distilled and turbo models

A distilled model is trained to imitate the output of a many-step model in very few steps: often 4 to 8. They are useful and they are how draft passes become affordable.

What they cost you is real and worth naming:

• Reduced motion range. Distillation tends to pull toward the average trajectory. Turbo outputs are noticeably calmer, which sometimes flatters a shot and sometimes kills it.
• Weaker prompt adherence on unusual requests. Common instructions survive distillation. Odd ones get smoothed toward the common case.
• Guidance behaves differently. Most distilled models want a guidance scale near 1. Feeding them the value you use on the full model produces burnt contrast and frozen motion.

The workflow that follows from this: draft on the turbo variant to test whether a motion idea works at all, then do the final on the full model with the same seed and still. The draft is not a preview of the final: it is a cheap answer to a different question.

Why the same settings give different pictures

Two runs with everything identical except the seed produce different clips because they start from different noise. That is expected. But two runs that you believe are identical can also differ, because the effective inputs include model version, precision, batch size and the specific GPU kernels used. The next lesson deals with that properly.

The free path

Every point above can be tested on a free Hugging Face Space or a Colab session, because they are all about settings rather than about scale. Fix the still and the seed, move one number, compare. An afternoon of that teaches you more about your chosen model than any comparison video, and it costs a queue rather than a balance.

Today: fix a seed and a still, generate at four step counts, and find the point on your own model where the improvement stops being visible. That number is now your production default.