VidFlow is one straight line: pick a channel, pick a title, write a script, lock the look, record the voice, generate the shots, render the video, ship it. The interface groups that line into five tiles — Ideation, Script, Visual Bible, Voiceover, Production — but underneath there are eight working stages plus a couple of bookends.
Pick a channel and a topic. Ideation is the first tile. You point VidFlow at a YouTube channel (yours, or one you're studying), it pulls the channel's recent uploads, scores them against an outlier model, and surfaces the topics that punched above their channel's average weight. The LLM then proposes a slate of titles tied to those outliers. You pick one. Nothing downstream runs until you do.
Write the script. Script is the longest editing stage and the one with the most knobs. The outline gets drafted first — the chapter count scales with your target length, from roughly three to five chapters for a short 5-minute video up to eighteen for a 30-minute one. Each chapter is then expanded into prose. The hook (your first 10 seconds) lives as a separate, rerollable beat so you can throw away ten openers without rewriting the rest. Per-chapter regenerate keeps the structure and replaces one chapter at a time.
Lock the look. Visual Bible is where character consistency happens. The LLM extracts every named character and location from the script. Each character gets a portrait plus a multi-angle sheet generated by KIE. Each location gets a base reference plus variations for time-of-day and viewpoint. You approve the bible before anything else runs. That approval gate exists because a bad character portrait will haunt every scene that character appears in — easier to fix once than fix forty times.
Record the voice. Voiceover stitches narration from one of three TTS providers (ElevenLabs, Inworld, MiniMax) into per-chapter audio. AssemblyAI then takes the audio and produces word-level timestamps that the next stage relies on. If you've got named characters with their own dialogue, MultiVoice mode lets each one have a different voice ID.
Produce the video. Production is the most expensive stage. The shot planner reads the timestamped script and decomposes it into semantic beats (HOOK, CLAIM, EXAMPLE, TRANSITION, CONCLUSION). Each beat becomes one or more shots. Each shot generates an image first (KIE nano-banana-2), then a video clip from that image (Kling 2.6 by default). Shots run in parallel — by default an in-browser pool keeps several in flight at once — and you watch them land in real time. (There's also an opt-in BullMQ worker path, off by default, for running generation in a separate process.) Then you review each one, approve or kick it back, and render.
Credits, honestly. A 3:30 finished video costs around 220 credits end-to-end. Video clips are the biggest line item by far — each Kling clip is a flat ~14 cr regardless of length. Image generation is the cheaper bulk (dozens of calls at ~2 cr each); voiceover and rendering combined are usually under 20%. The /pricing page has the up-to-date numbers — they shift as vendor prices shift.
Two modes, pick one. Quick mode is the 5-minute taste — pick a topic, hit one button, get a video. Most defaults locked, ideal for trying the product. Director mode is everything above, with every panel exposed. Most actual videos get made in Director mode. You choose the mode when you create the project, and it's fixed for that project — so if you want full control, start in Director mode rather than expecting to flip a Quick project over later.
When something fails. Failures usually land at the shot generation step (vendor content-policy reject, or a transient KIE timeout). Each failed shot keeps a lastError string with the reason and the Production card shows it inline under the FAILED badge. Click Generate on that shot to retry. Credits for shots that failed before producing any usable artifact are refunded automatically.
Start here, ship a project, the rest of the docs make a lot more sense after you've been through it once.