A failed video often leaves behind only a verdict: wrong, unconvincing, try again. Reopen the task days later and the same problem has to be diagnosed from scratch.
This library keeps different information: the time of the error, the constraint that failed, and the smallest useful change to try next. Its seven original categories were compiled on July 27, 2026, from this project’s production and research notes. They are neither universal model limits nor a collection of magic prompts.
Record symptoms before explaining them
“The characters never make contact” is an observation that can be checked in the footage. “The model did not maintain their spatial relationship” is an explanation. Mixing the two turns an old interpretation into an apparent fact when the model or shot changes.
Each entry keeps three layers: what happened, the current explanation, and the test still needed. Without a controlled comparison, it should not claim a particular word caused the problem. A completed generation is also a completed task, not an accepted film.
Seven failures worth keeping separately
| Category | Visible symptom | First useful test |
|---|---|---|
| Contact and collision | A reaction occurs while the weapon is still distant; feet float after landing | Test approach, contact frame, and reaction in a short segment |
| Multiple-reference contamination | One character inherits another’s helmet; a turn sheet becomes several characters | Reduce inputs and assign each reference a specific job |
| Premature evasion and idle pauses | A dodge precedes the threat; an approach ends in an unexplained stop | Define preparation, commitment, weight transfer, and recovery |
| Under-specified camera motion | Only the actor moves; a push becomes a zoom | State start, path, occlusion, and end; inspect parallax |
| Full-image redraw and style drift | A helmet fix also changes the face, material, and rendering style | Return to the approved master and change one region |
| Failed generator self-review | A pass report coexists with wrong action, identity, or timing | Inspect the final file independently |
| Explanatory-graphics failures | Vertical titles overflow; curves, labels, or initial frames are wrong | Check layout space, geometry, and exported frames |
Different categories require different responses. A contact-distance error needs a testable spatial correction; “more cinematic” does not supply one. A wrong mathematical curve is not redeemed by an attractive overall mood.
Actions need causal order
An action can fail through sequencing rather than an individual pose. A convincing dodge happens before the threat begins; an attacker reaches the target, stops, then strikes. Simply forbidding an early dodge may instead produce a static wait.
A better short test describes a sequence: continuous movement, a visible threat preparation, commitment to the attack, then evasion and recovery. An attack can be checked through final approach steps, planting, rotation, strike, and follow-through. Natural breathing and pauses should remain possible.
The contact-boundary record asks a narrower question: does the image actually support the claim that a hit occurred? Sound, debris, and reactions help communicate an event, but they cannot repair contact geometry explicitly required by the brief. If precise collision is central to acceptance, redesign the shot or consider an engine workflow with controllable state.
Give every reference a job
Before supplying several images, decide whether each controls identity, clothing, pose, setting, or material style. If two references compete over appearance, an output that mixes their features is unsurprising.
The project’s starting point was a small set with clear responsibilities. Face and full-body constraints were handled separately where needed. A multi-view design sheet may be useful to a human without being a suitable direct video input: its layout can become a group of characters.
For a shared frame, approve the subjects individually, then inspect their combined scale, sightlines, and positions. Two acceptable single-character images do not automatically make an acceptable composition.
Local edits need a way back
If a helmet correction turns a painterly image into plastic-looking CG, the edit has exceeded its intended scope. Continuing from that version makes the approved identity and materials harder to recover.
Keep the approved master, specify the editable region and invariants, and compare face, silhouette, material scale, and lighting. If one joint chain needs work, start with that chain rather than asking for a complete anatomical redraw.
The library used two budget rules: redesign the shot after three consecutive rounds of the same unresolved contact error, and return to the master after two local edits introduce new identity or global-style errors. These are project stopping conditions, not measured model success rates.
Explanatory video also needs correctness checks
A horizontal composition shrunk into a vertical frame may no longer fit its title. An intact face can still be obscured by captions. An elegant curve can incorrectly turn a parallel shift into a change of slope.
For these graphics, prefer inspectable SVG, HTML, or code. Calculate text width, the area reserved for a face, coordinates, and label placement explicitly. After the browser preview works, inspect the first frame, transitions, and key diagram frames in the final export. Initial timeline state can still produce a flash.
The aim is to give each risk an appropriate checkpoint: parallax for camera movement, distance and timing for contact, geometry and text for diagrams, synchronization for sound and image.
Make an entry useful to the next attempt
A reusable record should supply:
- The goal and invariants.
- Model and version, reference responsibilities, and original instructions.
- Original output, timecodes, and necessary frame captures.
- Observations written separately from causal hypotheses.
- The single change attempted, its outcome, and the stopping condition.
A failure entry earns its place by reducing blind retries. If it preserves only “try another prompt,” the useful judgment has not yet been recorded.