Review identity and product drift in image-to-video outputs

Run one bounded drift review before regenerating

Treat image-to-video drift review as a comparison against an approved source, not as a general impression of whether the clip looks polished. Freeze the source still, downloaded output, exact prompt, returned model label, ratio, observed duration, and generation ID before editing either asset. Create a review copy of the clip, but keep the delivered file unchanged. The finish condition is a recorded pass, reject, or escalate decision tied to specific frames.

Use five steps: verify the reference, extract a time-indexed frame set, compare protected details, inspect motion between samples, and record the decision. FFmpeg's documentation shows that `ffmpeg -i output.mp4 -fps_mode cfr -r 1 -f image2 'review/frame-%03d.png'` extracts one frame per second into a numbered sequence. It also explains that output `-r` duplicates or drops frames to reach the requested constant rate. That sample is an index, not complete evidence: scrub every frame around a suspected change and add the first, last, and event-specific frames to the set.

Reference gate

One approved still and its asset identifierIf reviewers disagree about the source of truth, stop; a clip cannot pass an undefined comparison.

Sampling gate

Opening, closing, one-per-second, and event framesAdd adjacent frames around fast motion, occlusion, camera turns, and any observed discontinuity.

Decision gate

Pass, reject, or escalate with frame numbersDo not replace a precise observation with a score that has no project-specific threshold.

Keep review copies inside the approved privacy boundary

A local frame-extraction and comparison pass does not require uploading the source or clip to another recognition, scoring, or vision service. Store review frames in the same approved workspace as the generation record, limit access to the assigned reviewers, and remove disposable review copies under the project's retention policy. Do not paste a confidential face, package, or unreleased design into an unrelated analysis tool merely to obtain an automated similarity score.

This local review step does not undo the generation-time cloud disclosure. OfflineCreator Studio's current privacy policy says prompts, settings, and source media submitted through MCP, CLI, API, or web flows are processed alike; generation requests send the prompt, settings, and source media to fal.ai. It says generation history and media remain until the user deletes them or closes the account, subject to short backup windows, while temporary uploads and failed-job artifacts are typically deleted within seven days. It directs users to LocalForge when work cannot leave the device. Apply the stricter project rule when contracts, consent, or embargoes require it.

Local review set

Reference, clip, extracted frames, and decision recordKeep these within the already approved project storage and access boundary.

No secondary upload

Do not add a new cloud processor by defaultEscalate before using face recognition, OCR, similarity APIs, or shared public review links.

Cleanup

Delete disposable derivatives on schedulePreserve the approved source, delivered clip, and decision evidence required by the project record.

Define the specimen before judging the motion

Write a specimen card for this exact run. Reference fields should name the approved source version, protected regions, allowed motion, permitted crop change, and the reviewer who approved those constraints. Output fields should record the actual model label, prompt revision, dimensions, duration, and file hash or immutable asset ID. The card prevents a later prompt, retouched still, or re-encoded clip from silently replacing the material under review.

Separate documented limitations from observations. OfflineCreator's current terms say model availability, behavior, speed, pricing, and output quality can change, do not guarantee that a generation will be accurate, unique, or suitable, and make the user responsible for reviewing outputs before use. Those statements justify a human review gate; they do not predict that a face, label, or product will drift in this clip. Record only what is visible in the compared frames and avoid turning one rejected run into a defect-rate claim.

Protected

Identity, geometry, approved text, color, and distinctive marksName only details that the source and project brief require to remain stable.

Allowed

Specified subject, camera, lighting, or environmental motionA deliberate change is not drift when it stays within the approved motion brief.

Observed

Frame-specific difference without a causal claimUse language such as “label edge changes at frame 73,” not “the model always fails on labels.”

Convert the motion prompt into review questions

Annotate the prompt before opening the clip. For `slow clockwise orbit; preserve bottle silhouette, cap geometry, label position, and warm-black background; one reflection crosses the glass; settle on the original view`, create one question per instruction. Does the camera travel in the requested direction? Does the silhouette remain compatible with the source as perspective changes? Do cap proportions and label placement stay stable? Is the moving reflection the only intended lighting change? Does the final view settle rather than morph?

Add source-only checks that the prompt may not mention. For a person, compare face shape, hairline, skin marks, eyewear, wardrobe seams, accessories, and hand count. For a product, compare silhouette, openings, fasteners, logo geometry, exact readable copy, package count, contact points, reflections, and material boundaries. For the scene, compare horizon, shadows, background objects, occlusion order, and edge behavior. These are editorial prompts for observation, not empirically validated defect classes.

Instruction

What was meant to moveCheck direction, timing, and end state without assuming that motion excuses identity changes.

Invariant

What approval required to remain stableTie each invariant to a visible source feature and a project requirement.

Tolerance

What perspective or lighting may legitimately changeEscalate ambiguous cases instead of labeling all pixel movement as a defect.

Build a source-to-frame contact sheet that exposes differences

Place the approved still in the first cell, then arrange the clip's opening frame, numbered samples, flagged adjacent frames, and final frame in chronological order. Use one scale for the full-frame row. Add separate crops for the face, label, logo, hands, product edges, or other protected regions, but label every crop as a review enlargement so it is not mistaken for a full output. Put timestamps and original frame numbers under each sample.

ImageMagick's official `compare` tool can create a visual difference image and report mathematical measures such as RMSE. Its documentation says a direct comparison starts pixel by pixel at the page offset; when image sizes differ, unmatched areas are treated as virtual pixels and can affect the metric. Aligning a source crop to an output frame is therefore a reviewer-controlled preprocessing decision. Use a difference layer to locate changed regions, not to declare identity preserved, text correct, or a clip approved. Perspective, intended motion, resampling, compression, and lighting can all create pixel changes that still require visual interpretation.

Overview row

Source, start, timed samples, endKeep chronological order and expose the entire frame rather than showing only favorable crops.

Protected-detail row

Same region at labeled timestampsUse consistent crop bounds where possible and disclose any alignment or resizing.

Difference aid

Optional visual map plus named methodNever let a pixel metric replace the project's semantic acceptance checklist.

Choose the next guide from the rejection reason

If the protected source was never approved or the animation path is still unclear, return to the approved-image workflow and establish the input record first. If the clip fails because the prompt asks for conflicting movement or needlessly recreates the scene, move to the motion-prompt guide and revise one instruction at a time. Use the workflow directory when review changes the deliverable rather than the next prompt. Keep the current contact sheet and rejection note beside the next run so reviewers can tell whether the targeted defect improved without hiding a new one.

Keep drift findings separate from provenance claims

This guide owns the narrow task “image to video drift review”: extract frames, compare them with one approved source, document protected-detail changes, and reach a human decision. It does not own motion-prompt writing, source-rights approval, model selection, or general MCP generation setup. The method is an editorial checklist, not a benchmark, automated identity test, defect taxonomy, model ranking, or claim that a particular prompt prevents drift.

Content Credentials may add useful history and integrity evidence. C2PA defines them as cryptographically bound provenance structures that can record origin, modifications, and AI use, but states that provenance does not judge whether the content is true, can be incomplete, and can be removed. Inspect a credential when present and preserve it with the record; do not treat its presence as proof that the face, product, text, or scene matches the approved source. This research pass generated no clip and provides no matched source/output artifact, measured defect frequency, or relevance-qualified community account, so the contact-sheet section is a reproducible template rather than output evidence.