Observability ads are dashboards; lock the screenshot

Observability ads are dashboards. The graph is the product. Image-to-video after the graph is true. Text-to-video of "a beautiful Grafana wall" is a cousin dashboard the buyer cannot log into.

If the still is wrong, every second of video is waste. That is the same routing rule as lock the still before you spend on motion, applied to a UI the SRE will pause on. A fake latency curve is not a style. It is a false claim.

The graph is the SKU

A staff engineer watching a paid social ad for an observability product is not looking for cinematic smoke. They are looking for whether the axes, the legend, and the alert copy look like the product you will actually demo. If the line goes the wrong way, they notice. If the panel chrome is invented, they notice. If a generated "P99" number never existed in your app, that is a listing problem wearing a motion pass.

So the production order is not "prompt a dashboard, then pick the prettiest take." It is:

1. Open the real product. Put the panel in the state the ad is allowed to claim.
2. Screenshot it. Crop to the ratio you will ship. That PNG is the still.
3. Approve the still the way you would approve a slide in a sales deck — numbers, labels, time range, redaction.
4. Only then buy motion.

Identity here is not a face. Identity is the panel. A retry of a text prompt is a new casting session for a dashboard you never locked.

Motion is camera, not a second typesetter

Image-to-video treats the upload as frame one. Composition, chrome, and the line on the chart are already decided. Your prompt describes what changes: a slow push, a cursor, a gentle drift across a wide layout. Leave the description of the graph out of the paragraph. Repeating the metrics in text is how you invite the model to renegotiate them.

Some models accept a last frame as well as a first. Use that when the shot has a destination you also screenshotted — the same panel after the incident resolves, the same service map with one node selected. Do not invent the "after" in a prompt. Capture it.

Length is a dial. Most image-to-video rows meter per second. A five-second hold on a true graph beats a fifteen-second wander that starts morphing tick labels. If the take drifts, shorten it. Do not "fix" a bad I2V by throwing the last frame at a text-to-video row.

What you may invent, and what you may not

Invent atmosphere around the locked panel: room light, a shallow desk, a phone bezel if the still is already the UI. Invent a camera move.

Do not invent:

• Series, thresholds, or error budgets that are not in the screenshot.
• A competitor's chrome restyled as yours.
• An outage you did not have, drawn as a spike, because it looked dramatic.
• Tiny axis type. On-screen text a video model draws has to hold every frame. Axis labels are below that floor. If the number matters, it is already in the PNG, or it is a caption you burn after.

Captions, a spoken line, a CTA — those are editor jobs on top of the take. They are not a reason to regenerate the dashboard.

If you do not have a true screenshot yet, you are not in an ad. You are in a mock. Stay in stills until someone can point at the panel and say that one would survive a demo.

FAQ

Can I generate a prettier version of our dashboard and then animate that?

Only if the prettier version is still true — same series, same labels, same product chrome — and someone signed it. A restyled lie is still a lie. Prefer the real screenshot.

Why not text-to-video if the model is good at UI now?

T2V invents the panel and the motion. That is two jobs, and the expensive one is how you discover the axis is wrong. I2V is the contract that the frame you uploaded is frame one.

Do I need a last frame?

Only when the shot must land on a second approved screenshot. A hero panel with a slow push is the usual job. A last frame is control, not a default.

What if the dashboard is too dense to read at 9:16?

Crop before you animate. The still sets the aspect. A 9:16 upload stays 9:16. Do not generate a wide wall and hope the model keeps the one series that matters.