AI video production is a chain of decisions: invent a shot, edit a story, then deliver a usable asset. Compare three documented workflows and prepare your own footage before spending generation credits.
This is a documentation-based comparison, checked October 5, 2026. We did not benchmark these services, buy plans or generate paid outputs. Model availability, quotas and terms change; verify your chosen workflow in the linked official documentation.
Choose for: reference-driven generative shots and model-specific video workflows.
Runway’s API documentation distinguishes text-to-video, image-to-video and video-based models. Supported aspect ratios and asset encodings depend on the selected model. A centered automatic crop can change a reference composition.
Producer question: which exact model, input route and crop behavior will preserve your intended framing?
Official input documentationChoose for: generating clips from text and working with explicit camera/style controls in Firefly.
The Generate Video documentation walks through model selection and generation controls. Firefly also offers partner models; record the actual selected model rather than assuming every output shares identical terms.
Producer question: are you selecting Adobe’s video model or a partner model, and which usage conditions apply?
Official generation workflowChoose for: transcript-led editing of recorded speech and subtitle handoff.
Descript describes editing video by editing the transcript. Its subtitle export workflow produces sidecar subtitle files. A transcript edit and an invented visual shot solve different production problems.
Producer question: do you need editable captions, burned-in captions, or both for the final platform?
Official transcript editing explanation · Official subtitle exportNo single service wins every stage. For an interview, accurate edits and caption review may matter most. For a missing establishing shot, references and controlled generation matter. Test one representative shot before committing a whole sequence.
Choose a browser-playable video, select 4–24 evenly spaced frames and build your sheet. Frame timestamps are approximate browser seeks, not certified frame-accurate edit points.
Load a video or try the demo. Nothing is uploaded.
Single-thread ffmpeg.wasm runs on your device. Its engine downloads only when you request an export (about 31 MB). The output is a WebM/VP8 video with Vorbis audio when the source has sound. Processing can be slower than playback.
Load a video above before exporting.
Compatibility: MP4/H.264 is a practical starting point; codec support varies by browser. HEVC, ProRes and unusual containers may require conversion first. Large sources can strain device memory. If extraction fails, try a shorter lower-resolution copy. The demo records an original animated canvas with MediaRecorder when available.
that carries your decisions.Describe the shot, retain source measurements and export a JSON or Markdown production brief. These fields are your decisions, not AI-generated recommendations. This lab makes no external generation requests.
No project data is saved between visits. Export a copy to retain your decisions.
Keep source footage, generated candidates and approved edits separately identified.
Inspect the uploaded video’s duration and dimensions, then extract frames to discuss composition. Browser seeks are approximate; inspect final edits in your native editor when frame accuracy is required. A reference should clarify the subject and motion rather than merely resemble the intended color palette.
State the subject, movement, camera behavior and intended end state. Separate unrelated events into different shots. This is an editorial suggestion, not a guarantee that a model will follow every instruction.
Look for identity drift, temporal discontinuities, unreadable text, unwanted crops and incorrect actions. Scrubbing frames can expose a defect but does not replace reviewing the full sequence with sound. Record rejected candidates and the specific reason for rejection.
The prep lab emits a VP8 WebM review copy, not an archival master or a guaranteed platform-compatible delivery. It rescales pixels with FFmpeg; it does not recover detail with AI upscaling, stabilize a shot or certify broadcast compliance. Verify your destination’s format before a final delivery.
“What changes between the opening and the final frame?”
Product descriptions above link to their official documentation. Local frame extraction uses the browser’s video decoder and Canvas. Local trimming uses pinned ffmpeg.wasm 0.12.15 with single-thread core 0.12.10, GPL licensing included. Processing may consume substantial memory and time on mobile.
FFmpeg scale filter · ffmpeg.wasm overview · Core license · Wrapper license
Your selected files stay in this browser. This page downloads its runtime assets, fonts and motion libraries; it does not submit footage to any AI service. References and exported briefs may contain personal information: review them before sharing.
Prepare your next shot