We needed a four-second product video: an industrial sleeve shrinking down onto a metal fitting. Simple physics, one subject, no dialogue. It failed on two different paid platforms.
The obvious conclusion was that AI video was not there yet for real physical products. That conclusion was wrong, and it cost us two subscriptions to find out.
So we ran a controlled test instead of buying a third plan. Same model. Same prompt, word for word. One variable changed: the seed image.
With the original seed, the model re-composed the frame, invented a background that was never requested, and the physics test was void — the product was already in its finished state, so there was nothing left to animate. With the corrected seed, the framing held exactly and the material closed down onto the fitting the way it does in reality. No warping, no sagging, no drift.
Same model, same prompt, opposite outcome. The control was never in the prompt.
Why this contradicts what everyone is publishing
The current wave of AI video content is lists of slash codes — /cinematic, /dollyin, /goldenhour — usually presented as secret or official commands, often numbering in the hundreds.
Two things are worth knowing before you memorize any of them.
First, they are not official commands. They are community shorthand for ordinary cinematography vocabulary. Google's own guidance is to describe the scene in natural language, including subject, action, environment and lighting. A slash prefix does not invoke a feature; the model is reading the word, not the symbol. That is not an argument against using them — shorthand is a genuinely efficient way to write a consistent prompt — but a list of 109 of them is a glossary, not a control panel.
Second, and more practically: some of these tools do not send your prompt to the model unchanged. We dropped one popular agent-driven interface from our evaluation for exactly this reason — it rewrites the prompt on the way through. That makes it fine for casual generation and useless as a test surface, because you cannot attribute a result to an input you did not control. If you are carefully tuning codes inside a tool that is editing your words, you are tuning something you cannot see.
Meanwhile the variable that actually decided our output was one nobody in that content was discussing.
The two rules that came out of it
1. Seed the state you want to animate from, not the finished product.
This sounds obvious written down and it is very easy to get wrong, because the instinct is to seed with your best image — the hero shot, the product looking its finest. But your best image is usually the end of the motion you want. Seed a sleeve already shrunk and there is no shrink left to perform, so the model does something else with the four seconds. Seed it loose, with a visible gap, and the motion has somewhere to go.
Ask what the last frame should be, then seed the first one.
2. Use a single subject on a real background.
Our failing seed was a multi-object composite on a transparent ground. That is close to an instruction to re-compose: transparency gives the model no spatial information, so it invents some, and multiple objects give it no clear subject, so it chooses one. A single subject photographed in a real environment removes both ambiguities. Framing then tends to hold, which is what we actually needed.
Both rules are about removing decisions from the model rather than adding instructions to it. That is the shift in thinking — most prompt advice is additive, and the wins here were subtractive.
The cost structure nobody mentions
Three things about generation economics that changed how we work, none of which appear in prompt guides:
- Image generation is roughly five times cheaper than video generation. Which means building the correct seed is the cheapest step in the pipeline, and iterating a seed image is far cheaper than iterating clips. We now spend deliberately more time before the first video generation.
- A failed generation still charges you. An errored run consumed a significant share of a test budget and returned nothing. Generation is metered on attempts, not on usable output, so a workflow that produces four throwaways per keeper costs five times what the pricing page implies.
- The in-app credit counter can be wrong. On one platform the remaining-generations indicator read low double digits when the true figure was three. Track spend against the account ledger, not the interface.
Put together: the real unit cost of AI video is not the advertised price per clip. It is the price per clip multiplied by your failure rate — and the failure rate is mostly determined by the seed.
What the codes are actually good for
To be fair to the format, shorthand does earn its place. It makes prompts consistent across a batch, which matters when you need six clips that look like they belong together. It is a useful vocabulary for people who have not worked with cameras, because it names things like rack focus that are otherwise hard to describe. And it makes a prompt reviewable at a glance.
What it does not do is give you control over composition and physics. Those are decided earlier, by what you hand the model.
Use the vocabulary. Just do not expect it to fix a seed that has already told the model to improvise. If you want a structured way to build the descriptive part of a prompt, our JSON prompt generator does that job for image and video models.
The honest limit
This was one product category, tested on a small number of models, over a defined evaluation period. Product physics on a single rigid subject is close to the friendliest case there is — we have not established that seeding governs equally for human motion, crowds, or long multi-shot sequences, and we would expect it to matter less as scenes get more complex and other variables dominate.
Model behavior also moves fast. A rule derived in one release can be softened by the next as models get better at inferring intent from imperfect inputs. Treat this as a finding with a date on it, not a law.
And the honest caveat about the cost figures: exact credit prices change constantly, so we have deliberately given ratios rather than numbers. The ratios have been stable; the prices have not.
Five questions before your next generation
- Does your seed show the beginning of the motion, or the end of it? If it is the end, you have nothing to animate.
- Is there exactly one subject, on a real background? Composites on transparency invite re-composition.
- Does your tool send your prompt through unchanged? If it rewrites, stop tuning wording — you are not controlling the input.
- What is your cost per usable clip, not per generation? Failures bill the same as successes.
- Have you iterated the seed before iterating the video? It is several times cheaper to fix the input.
We ran this evaluation because we produce video for client social channels and needed to know where the control actually was before spending client budget on it. If you want that applied to your own content production rather than your own credit balance, that is what our social team does.