GENERATIVE FILM SHOT CONSISTENCY CHECKLIST
The nine tells, arranged by the stage at which each has to be caught
- Format
- Checklist, PDF and on this page
- Length
- 27 checks
- Gate
- None
- Licence
- Use it, commercially
Nine ways a generative sequence gives itself away, arranged by the stage at which each one has to be caught. Applied before the render, all of them are free. Applied after, most of them are a regeneration.
Nothing on it is model-specific, and none of it stops being true when the models improve. Run it as written or cut it down to the gates your shot types actually need.
DOWNLOAD THE PDFNo email, no form. The whole thing is on this page too.
- 01Run section 1 before generating anything. It is a set of decisions, and it takes twenty minutes.
- 02Run section 2 on every frame or clip as it arrives. Pass or fail, not a discussion.
- 03Run section 3 on the assembled set. It catches the failure no per-frame test can.
1. Before you generate
Decisions, written down. Every one of these is free now and expensive later.
- Key light direction fixed and written as a clock position, not as "soft natural light"
- Colour temperature stated in kelvin
- One surface and one background material named
- Lens language chosen: focal length range, and shallow or deep depth of field, held across the set
- A prop rule written: what is allowed in frame, and what is never
- Shot-size pattern decided before the render, not discovered in the edit
- Master plate built and frozen for every product that must not change
- Trained identity built for any face appearing in more than nine assets
- Credit ceiling per asset agreed, and encoded so the run halts rather than warns
2. Per frame or clip
Pass or fail. A failure returns to the source file, never to a retoucher.
| GATE | THE TEST | FAIL MEANS |
|---|---|---|
| Type | Every printed word read at 100% against the artwork | Regenerate. Never retouch a label. |
| Silhouette | Overlay on the master plate at 40% opacity, inspect the outline | Regenerate from the plate with tighter conditioning |
| Identity | Compare against the identity sheet: jaw, eye spacing, apparent age | Dead shot. Regenerate. |
| Light | One key direction and one temperature, matched to the specification | Regenerate the outlier, not the set |
| Claims | No invented certification, award, ingredient or percentage | Kill the frame. No appeal. |
| Extremities | Count fingers. Check contact points and occlusion | Regenerate, or cut earlier |
| Physics | Gravity, contact, cloth settle, hair settle | Cut before the drift, or regenerate |
| Text in frame | Any readable word the model rendered | Recompose out of frame, or composite in post |
| Duration | Ends before extremity drift, not when the model stops | Trim to the stable opening segment |
3. On the assembled set
The failure no per-frame test catches
- Contact sheet at thumbnail scale: anything that jumps out gets regenerated even if it passed every gate
- First frame and last frame of every clip, side by side at full size, checked for drift
- Every clip scrubbed backwards at speed, which reveals drift the forward motion hides
- Grade applied to the sequence from one reference frame, after picture lock, not per clip
- Room tone under everything, foley on the two or three actions the eye lands on
- No cross-dissolve between two generated clips anywhere in the edit
- At least one deliberately quiet beat in the sequence
- Disclosure applied on the asset, and provenance metadata on export
If you only ever apply three things from this document: fix the key light direction, hold the face, and cut before the motion tell. Those three account for most of the perceived quality gap.
Free to use, copy, adapt and put your own name on. No attribution required, though a link back is appreciated. Published by TaleCrafters, talecrafters.studio.
- How a Master Plate Works in Synthetic Product ProductionThe single technique that separates a product set that holds from one that quietly falls apart by shot forty. It is not a prompt.
- How We Keep a Product Consistent Across 100 AI-Generated ShotsShot one is easy. Shot one hundred is the job. This is the method, including the numbers we plan against.
- Why Your AI Video Looks Cheap, and What Fixes ItAudiences cannot name what is wrong with most generative video. They can tell. Here are the nine things they are reacting to.
- Trained Identity vs LoRA vs Reference Image: Choosing a Consistency MethodThe right method depends almost entirely on how many assets the face has to survive. Below nine, one answer. Above nine, a different one.
- Temporal Coherence: Why AI Video Falls Apart After a Few SecondsThe limitation is real and it is not going away this year. The good news is that cinema solved this problem in 1918.
Using this on a live job?
If the checklist turns up something you would rather not run yourself, TaleCrafters does this for a living. Bring the brief and the deadline.