What Does an AI Short Film Actually Cost? Night Huntress Season One, Full Ledger

“AI video is cheap” is half true. When Night Huntress season one wrapped — eight episodes — we opened the books: the cash spend really was absurdly small, but only because the bulk of the ledger was never money. It was rebuilt versions and verification hours, plus one licensing bill that surfaced only after delivery. Here is the whole ledger, including the line items that were never recorded — where nothing was logged, this article says so instead of back-filling a plausible number.

ENVIRONMENT

Night Huntress season one, 8 episodes plus a compilation cut (16 shots × 10 seconds per episode, 1280×720, completed early 2026-10)

LEDGER SOURCES

the season’s production records and workflow document (locked 2026-10-02), the finished-version files for every episode, and the music licensing confirmation memo (2026-10-14)

GENERATION

image generation (stills) plus first/last-frame image-to-video, run through the studio’s existing AI toolchain — not billed per piece

VERIFICATION

frame extraction and human review of every segment (a fixed workflow step, including contact sheets at 2-second intervals)

Start with the denominator: how much was actually made?

A ledger needs its denominator first. Season one’s output, all from the production records:

“At least” is the honest boundary: per-shot regeneration counts (how many attempts each shot took before it locked) were never tallied shot by shot, and this article won’t reconstruct them after the fact. What can be counted from files is whole-episode re-renders — that’s the next table.

The version ledger: why was episode 3 rebuilt seven times?

The number of finished versions per episode is the most honest redo metric we have — every version is a complete re-assembly and re-render, and every one was triggered by something verification caught:

Nineteen episode versions plus two compilation versions: 21 full renders. Episode 3 alone accounts for seven of them — it was the season’s most expensive tuition. The face-reference workflow (article 1 on this site), the no-indoor-rain rule, and the equal-power crossfade rule for music all came out of this episode’s repeated failures. In other words, redos aren’t waste; they’re how a workflow gets bought. The later episodes could ship in one or two versions because episode 3 had already paid for the rules.

The time ledger: what does verifying one episode take?

Generation is cheap; verification is not. Every episode runs a fixed procedure: extract frames from every segment, build contact sheets, and review them by eye, shot by shot — checking whether the face drifted, whether rain fell somewhere it shouldn’t, whether costume pieces and props match the bible. As documented in article 1, that runs about 1 hour per episode. Across eight episodes, frame-checking alone is roughly 8 hours — and that excludes:

• The face reference itself: a one-time job of about 20 minutes (article 1).
• Storyboards and prompts: every episode starts as a 16-shot board — scene, characters, action, one narration line per shot — before anything is generated.
• Music handling: one theme looped across each episode, seams joined with 1.5-second equal-power crossfades; rain ambience only on exterior shots. All human judgment and listening time.

There is no per-task timesheet for the rest (another unlogged item); the anchor we can stand behind is the verification figure: about 1 hour per episode. When someone tells you an AI short costs only its generation fees, ask them who does the verifying.

The cash ledger: what was zero, and what is still open?

Line by line:

That last line is the reason this article exists.

The hidden cost: the score’s licensing surfaced after delivery

The season’s theme (Memory of the Blade) and episode scores were generated with ElevenLabs’ music feature — and at generation time, nobody thought to read the licensing fine print. On 2026-10-14, preparing a festival-platform submission (SinemAI), we read ElevenLabs’ Terms of Use and its music-specific terms end to end and wrote up a confirmation memo. The findings:

• General commercial use: paid users may use outputs commercially and retain rights in what they generate.
• But the music terms carve out film: except for the top Enterprise Music tier, every self-serve plan’s music commercial rights exclude film, TV, radio and large-scale games. Streaming distribution is fine from Creator up; “film use” is the part outside the fence.
• Night Huntress is a film work headed for a streaming platform — squarely inside the exception.

The memo lists three options: A. obtain written licensing confirmation from ElevenLabs (an inquiry letter is drafted; lowest cost); B. upgrade to Enterprise Music (custom quote, amount unknown; its terms apply upgrades retroactively to existing outputs); C. replace the score in the submission version (the most work; a replacement track list is already prepared). As of this article’s fact-check date, the bill is still open. The lesson generalizes: AI generation’s “cheap” sometimes means the licensing cost is deferred — deferred to the moment of submission, commercial release or distribution. Even the subscription tier in effect at generation time wasn’t archived (the memo lists it as outstanding), which is its own tuition: our project files now treat “plan tier and terms version at generation time” as a required field.

Failures and limits: what this ledger is missing

House rule — this section is never skipped:

• Per-shot regeneration counts were never tallied. We can count whole-episode versions (21) but not how many attempts each of the 128 shots took. Total generation volume is therefore a floor, not a figure.
• Subscriptions were never allocated per project. Generation ran on the studio’s existing toolchain, and its cost was never split out for Night Huntress — so this article prints no “total cash cost”. Printing one would mean inventing it.
• Labor has anchors, not timesheets. Verification at about 1 hour per episode is workflow-backed; storyboarding, prompting and mixing hours were not logged per task.
• The licensing bill is open. The score’s film licensing (which of options A/B/C, and the eventual amount) is undecided; when it closes, this ledger gets a follow-up post.

Sources and further reading

• Night Huntress video production workflow v2 (locked 2026-10-02): the 16 shots × 10 seconds spec, the verification procedure (frame extraction and contact sheets at 2-second intervals), and the music and rain rules are from here.
• The season’s finished-version files (DC Studio production files): the episode version counts (19) and compilation versions (2) were counted file by file.
• Night Huntress ElevenLabs music commercial-rights confirmation memo (2026-10-14): the terms check (ElevenLabs Terms of Use and Music Terms, official pages updated 2026-10-09), the film exception and the three resolution options are from here.
• Article 1 on this site, “How Night Huntress Kept the Same Face for 8 Episodes”: the face-reference workflow and the verification time (about 1 hour per episode) are from there.

FAQ