The companion episode — the film plays first, then she hands you the pipeline, failures itemized.

We made a short film: our presenter gets yanked out of her studio by a helicopter rope ladder, is chased by winged monsters, climbs to safety, discovers the cockpit is empty — and then the camera pulls back to reveal the whole thing is a film set. Total cost: $16, including every re-roll and rejected take. Here’s the pipeline.

1. Storyboard on paper first

One still image per shot. Generate storyboard plates in batches (we do 8 takes per shot and pick), with your character LoRA on every frame — yes, even wides. Fewest characters and locations wins; our 11-shot film has one character and three locations.

2. The three directing rules

Learned expensively, documented fully here:

  1. One beat per clip. One action per 5-second generation. Two actions = teleporting limbs.
  2. Natural speed, short clips. Long slow shots are where AI gets caught — objects drift, our dragons literally shrank. Fast cuts are camouflage and better filmmaking.
  3. Describe the whole outfit — the animation reveals what the storyboard cropped.

3. Animate: cloud, local, or both

  • Cloud (Wan-class via Civitai (referral link — supports the channel at no extra cost)): fast, strong motion, per-clip cost in platform credits. Caveat: cloud moderation is a lottery — innocuous athletic shots get refused; reword and reseed, or…
  • Local (LTX-Video): slower per clip, zero cost, zero moderation, runs overnight. Our rule: cloud for complex motion, local for everything else and every cloud refusal.
  • A blocked clip is information, not a wall — two of our shots rendered at home after the cloud bounced them twice.

4. Sound is half the film

The score makes or breaks it — audiences forgive AI visuals long before AI audio. Everything we learned scoring this film, short version: name emotion + instruments, generate multiple takes, pick by ear, and consider a dedicated music model (Suno ended our three-failure streak in one take). One craft move that outperformed its cost: cut the music mid-note at your twist — silence is a punchline.

5. Assemble

Normalize clips → hard cuts (the pacing is the transition) → score under → freeze-frame title card. The whole edit is one script, which means film #2 edits itself. (No script chops? A consumer timeline editor like Filmora (referral link — supports the channel at no extra cost) does the same job by hand — the moves are identical: hard cuts, score under, title card.) Add a voice line where the edit needs an author (our “CUT!!” is a stock TTS voice), and run a moderation self-check before publishing: sample a frame every few seconds and review the strip like a hostile algorithm would.

The honest part

Almost every shot in our film is a first or second take. That’s the point — this is the floor, made in days for lunch money by one human, one AI, and one GPU. The systems only get better from here. The music-video pipeline is this guide’s sibling; between them you can shoot almost anything.