An image-to-video model does not start from a script. It starts from a single picture and then has to guess everything the picture does not show. Anything ambiguous in that picture becomes motion you did not ask for. The cheapest quality control in the whole workflow is therefore a slow, deliberate look at the first frame before generation.
This checklist is written for Pavo-style image-to-video and text-to-video shots, but it applies to any model that animates a still. Print it, or keep it open next to your editor. It takes about ninety seconds per frame.
The 11 checks
- Subject silhouette reads at thumbnail size. Shrink the still to 120 px wide. If you cannot name the subject, the model will not be able to hold it.
- Contact is believable. Feet on ground, object on surface, shadow attached to the base. Floating contact shadows are the first thing to wobble.
- Hands are all in or all out. A half-cut hand at the frame edge is an invitation to invent a sixth finger.
- There is one light direction. Two competing highlights mean the model will have to choose, and it will choose differently on different frames.
- No readable type anywhere. Labels, packaging, signage and shirt prints become garbled glyphs once motion starts.
- The background has one depth story. A mirror, a screen or a window with a view adds a second scene the model must keep consistent.
- Identity anchors are visible. For a person: both eyes, jawline, hairline and at least one ear. For a product: silhouette, neck, base and one distinctive edge.
- Negative space is planned. Decide now where the caption, logo or lower third will sit later, and keep that area empty.
- The horizon is level, or deliberately and obviously not. A tilted horizon that you did not intend will read as a camera mistake.
- There is exactly one plausible motion. Ask yourself what the subject would naturally do next. If there are three answers, the shot is overloaded.
- The still is worth exporting on its own. If you would not use it as a thumbnail, do not animate it.
Turn the check into a six-line prompt
Once the still passes, write the motion as six short lines. This is the format used in every preset on this site, and in the free prompt builder.
| Line | What it decides | Common mistake |
|---|---|---|
| SUBJECT | What exists in the frame | Listing things you hope will appear later |
| FRAME | Composition, light, negative space | Using mood words instead of positions |
| ACTION | One chronological movement | Three simultaneous actions |
| CAMERA | One move, or locked off | Cinematic camera with no defined path |
| FIXED | Anchors that must not drift | Forgetting the light direction |
| END ON | A held frame for cutting | No ending instruction at all |
What to do when the still fails the check
Do not compensate in the prompt. A weak still plus a longer prompt produces a longer failure. Regenerate or edit the still: crop out the half-hand, remove the label, flatten the lighting, simplify the pose. One clean still is worth more than five clever adjectives.
When you are working from a photo you own, a small local edit is legitimate and faster than a re-render: mask the signage, straighten the horizon, clone out the extra reflection. Keep a copy of the approved still with a short name such as shot-03-approved.png. You will need it again for the second shot in the sequence.
A 90-second routine
- Open the still at 100% and check hands, contact, type and reflections.
- Shrink to thumbnail size and check the silhouette.
- Write the six lines, then read the FIXED line twice.
- Generate once. Watch at normal speed, then at 25% speed.
- Log the one thing that failed. Change one line. Generate again.
The rest of this site builds on that routine: twelve shot presets that already pass the check, a failure-to-fix table for when they do not, and guides on camera moves, character consistency and delivery.