What a generated 3D animation actually contains
The output is skeletal motion: joint rotations and translations over time, written into a portable file. It contains no geometry, no texture, no lighting, no camera, and no rendered frames, which is why it imports cleanly into a scene you already art-directed. This is the difference between animation data and a video. One is a resource your pipeline can retime, blend, and reuse; the other is a finished image you can only watch.
FBX, JSON, and the optional BVH step
Choose FBX when the next stop is a DCC application or a game engine, because most importers accept it and most retargeting tools expect it. Choose JSON when a technical artist wants to read channels, frame counts, or rotation values directly, or when a custom exporter needs structured input. Some accounts also see a newer beta conversion step that produces BVH, which suits older motion tools. None of these files is a certified production asset; all of them are drafts that need a review pass.
Tuning duration, guidance, and seed
A single clip can run from half a second to twelve seconds, so plan sequences as several overlapping drafts rather than one enormous take. Guidance decides how literally the movement follows the description, and the seed fixes a run so a colleague can reproduce it. When you compare two prompts, change one variable at a time and hold duration, mode, and seed steady. That discipline is what turns a folder of experiments into a decision, because a result you cannot reproduce is only a lucky preview.
Prompt structure that animates
Describe one humanoid, one chain of events, a direction, a tempo, and a final pose. Walk, stop, turn left, and settle is a complete request; make the character do something interesting is not. Keep it under about sixty words and leave out camera language, weather, emotion arcs, and second performers, because the model cannot stage them and will spend the clip guessing. If the first draft drifts, tighten the beat list rather than adding more adjectives. Short, physical sentences give the clearest timing to inspect.
Fitting a clip into an existing scene
Check frame rate, unit scale, and forward axis before the first real import, and confirm whether travel belongs on the clip or on a controller. Long clips are harder to blend than short ones, so split a sequence at natural pauses and cut on poses that resemble each other. Watch contact: a foot that slides or a hip that drifts will read louder than any rotation error. Finally, review the motion at the distance the audience will see it. Detail that impresses in a viewport often disappears in a wide shot.
The parts a person still owns
Physics, contact with moving scenery, loop closure, comedy, and performance remain human work, and a draft cannot answer them by itself. Treat every download as a proposal and record what you kept, what you repaired, and how long the repair took. That log is the only honest way to know whether the tool saves studio time on your kind of shot, because a cheap draft that needs three hours of cleanup is not cheap. Measure, then decide whether a wider run is worth the credits.