The boring file checks that save an expensive render.

A rejected file is not a creative problem

An expensive video job can fail before the model has done anything interesting. The uploaded file is too narrow. The link has expired. The file contains a rotation flag that the preview handles and the processing service does not. A twenty-second export is fractionally longer than twenty seconds. None of those problems gets better when you rewrite the prompt.

This is why a repeatable video workflow starts with an input contract: a short, explicit statement of what this particular model route accepts and what each supplied asset is meant to control. It is less exciting than a new generation model. It is also much cheaper than discovering the same constraint at the end of every upload.

There is no universal “AI-video-ready” format. Limits belong to an endpoint and version, sometimes to a mode within that endpoint. A reference-video editor, a motion-transfer tool, and an image-to-video model can share a product name while accepting different inputs. Save the actual requirements with the job. Do not rely on a screenshot from another mode.

The workflow below is a proposed production method. The numbers are an illustrative configuration, not a claim about the current limits of any named provider.

Inspect the media you have, not the filename

Keep an untouched source, then inspect the working copy. Record its encoded width and height, displayed orientation, duration, frame rate, video codec, audio streams, and file size. A filename ending in “720p” is not evidence that the encoded video meets a minimum-width rule.

The official ffprobe tool can report container and stream information in a machine-readable form. That makes it suitable for checking the actual file and saving a small report alongside it. A basic inspection can be as simple as ffprobe -v error -show_format -show_streams -of json input.mp4. The report is an input to a decision, not a guarantee that a remote service will accept every stream it describes. ^[1]

Suppose the source is a vertical 540 × 960 video and the chosen endpoint requires a width of at least 720 pixels. Scaling it to 720 × 1280 preserves its 9:16 aspect ratio and increases each dimension by one third. Stretching it to 1280 × 720 would satisfy a width check while ruining the picture. Cropping it without checking the composition might remove a hand, a product, or the top of the character’s head.

That resize creates more pixels, not new evidence about the face. A blurry source remains a blurry source at a larger size. Use the cleanest original available, and avoid a chain of downloads and recompressions before the model sees it.

Inspect the prepared file again. That second report catches a surprisingly useful class of mistakes: an export preset that was not applied, an unintended audio track, or an output that looks upright only because the player interprets metadata differently.

Give every reference one clear job

Write a small asset list before submission. The source video controls movement and timing. The character reference controls identity. A final-frame reference, when the selected mode actually supports one, controls the requested ending composition. The audio master supplies the original soundtrack. Those are separate responsibilities.

Do not describe a generic image reference as a guaranteed first-frame constraint. The interface may accept both, but they are different controls. Check which fields the endpoint exposes and how it describes them. If an edit model accepts a video and identity images, use those inputs as documented. Adding an unsupported end-frame parameter does not make the request more precise.

Reference quality matters more than reference quantity. Choose a face image with useful detail, ordinary proportions, and lighting that lets a reviewer see the person. A dramatically lit portrait can be excellent artwork and a poor identity anchor. Several contradictory reference faces can make the job less constrained rather than more constrained.

Save the exact submitted assets under stable names and record their content hashes. Also record the model route, settings, prompt, source duration, and any transformations. This record will not make a stochastic generation deterministic. It will tell you what actually changed between attempts, which is the minimum requirement for learning from them.

Split for a real limit, not out of habit

Prefer the full source when the chosen workflow accepts it. Every boundary you introduce creates another place to lose timing, identity, lighting continuity, or audio alignment. Splitting should solve a documented duration constraint, not substitute for understanding the tool.

Consider a hypothetical 73.4-second video normalized to a constant 30 frames per second. It contains 2,202 frames. If the selected endpoint accepts at most twenty seconds, use three 600-frame segments and one 402-frame segment. Their durations are 20, 20, 20, and 13.4 seconds. Together they cover the complete source without an intentional gap or overlap.

Frame counts are useful here because rounded timestamps can conceal an extra frame. They are not a reason to force every source to 30 fps: converting frame rate can duplicate or drop frames. Choose a consistent working rate deliberately, then record the conversion. A variable-frame-rate source needs particular care because a displayed average rate does not define every frame boundary.

FFmpeg distinguishes seeking while transcoding from seeking while copying encoded streams. Accurate seeking during transcoding can decode and discard material before the requested point; stream copying can retain material around an earlier seek point. That is why a fast copy operation is not automatically a frame-exact cut. ^[2]

After cutting, check each actual output duration against the endpoint’s limit. If its validator rejects a file at the exact boundary, prepare a smaller frame-aligned interval and update the manifest; do not silently discard the end of the source. Keep the segment order and covered frame ranges explicit.

Submit once, and keep the soundtrack separate

Before spending credits, confirm that the service can fetch the supplied media. A browser-accessible sharing page is not always a direct media URL. Expiring links, login requirements, incorrect content types, and incomplete uploads are transport problems, not model failures. Use the provider’s supported upload mechanism and confirm the upload has finished.

Then save the returned job identifier. A slow response does not necessarily mean the request failed. Check the existing job before submitting another paid generation. Keep a small ledger with the job status, estimated charge, actual output, and reviewer decision. The point is to distinguish “not accepted,” “still processing,” and “completed but unsuitable.” Those states require different next steps.

Preserve the untouched audio master throughout. If the provider returns a new soundtrack, do not let it quietly replace the approved original. Where the generated visual timing is unchanged, the original audio can be placed against the completed visual sequence. FFmpeg supports copying an audio stream without re-encoding when the output container and stream are compatible. ^[2]

Matching duration is necessary but insufficient. Check a visible speaking moment near the beginning, middle, and end. A model can return a clip of the right length while changing the timing of a gesture or mouth movement. Do not hide a synchronization problem by trimming the last second and calling the file finished.

Review the failure before buying another attempt

Separate four review outcomes: the file is invalid, the service cannot retrieve it, the generation violates the visual brief, or the delivered edit fails synchronization. “The video did not work” combines all four and encourages random retries.

For a visual failure, describe the first point where the result departs from the brief. A face change after an occlusion suggests a different next experiment from an incorrect outfit throughout. Keep other inputs stable while changing the reference or instruction you believe matters. Our identity-consistency review offers a way to inspect the full clip rather than approve a flattering thumbnail.

Finally, package the approved output with the source manifest, reference versions, audio master, and review notes. Someone should be able to reopen the job next month and know which source was processed, why it was resized, whether it was segmented, and which result was actually approved.

The creative brief still matters. So does the model. But when a production process repeatedly burns time on the same preventable input error, the next useful improvement is often a thirty-line validation check, not a more poetic prompt.