SCARLET BEAST — The Zero-Budget Production Plan

How to make a full-length animated feature on one borrowed-time GPU, a Windows workstation, free software, and people who love you.

Draft of September 26, 2026. Every number marked (est.) is an estimate to be replaced by a measurement in Phase 0. Numbers without that mark are either counts from the screenplay or things this setup has already done.

0. What we are making

The screenplay is complete. This document is how it becomes a film.

1. The strategy: always have a whole film

Most zero-budget features die the same way. Someone animates the first scene beautifully, then the second, and around scene nine life happens. What's left is ten gorgeous minutes of a film that never existed.

We do the opposite. At every point from week four on, a complete film exists, start to finish, that you could sit in a room and watch. It just gets better.

Why this works:

• The edit is where films are really made. An animatic tells you in week four that sequence 12 is eight minutes too long. Finding that out after animating it costs months.
• Every shot is always "done" at some level. If the project stops at v1, you still have a finished film.
• The work is upgradeable in any order. You can spend a good week on the finale and a tired week on transitions, and the film stays whole.
• It protects you. A creator with a finished animatic keeps going. A creator with 9% of a film doesn't.

Rule: no shot gets upgraded to v(n+1) until every shot exists at v(n). The one exception is test shots in Phase 0.

2. The look

Two worlds, one film

The screenplay runs on two layers that merge. The palette carries that before anyone says a word.

The film starts almost fully separated and ends fully merged. In the finale the Trenton lawns are the City. A grade LUT per act makes the merge measurable: Act 1 Earth scenes are ~70% desaturated, and by Act 3 they're at full saturation.

Two styles of acting

• Disney-Renaissance (drama, Realm, Christian, Mamá, the Scarlet Woman, Yeshua, Samael): weight, breath, held moments, full expressions, painted light. References: The Lion King, The Prince of Egypt, Hercules for the Realm's gold.
• Looney Tunes (the Seven Heads, Hiss & Growl, Rev. Buck Stallings, gag cutaways): bold outline, flat bright colour, squash-and-stretch, smears, smash cuts, holds on a reaction, fourth-wall looks. Snappy timing: the gag lives in frame counts, so these shots are animated on 2s with hard holds.

The rule that keeps them in one film: same line weight and the same scarlet. The Heads are funnier, not a different universe.

Model sheets (Phase 0 deliverable)

One sheet per principal: turnaround (front, 3/4, side, back), 6 expressions, 3 key poses, colour callouts for Earth and Realm lighting, and a size-comparison lineup.

Principals: Bryan/Christian (child 6, teen 15, adult; each needs its own sheet), Mamá, the Scarlet Woman, Samael, Yeshua (human form and Lamb/Lion-shadow form), the Beast (seven-head and crowned), each of the Seven Heads, Hiss, Growl, Nurse Grace, Chaplain Amos, Dr. Pell, Rosa, Deacon, Tiny, Rev. Stallings. Plus location sheets: the Bronx apartment, the Bronx rooftop, Trenton grounds and day room, Abaddon, Samael's court, the City.

The Trenton grounds already exist in 3D from the New Jersey, New Jerusalem Blender project (real OpenStreetMap footprints). That scene can supply camera layouts and backgrounds for every Trenton exterior and for the City's descent.

3. Character consistency

This is the hardest technical problem in AI-assisted animation. A character who changes face from shot to shot breaks the film. We solve it in three layers.

Layer 1 — a style LoRA for the whole film

Train one SDXL LoRA on the approved concept art, model sheets and paintovers, so every image comes out in this film's line and palette rather than generic "cartoon".

Layer 2 — a LoRA per character

The child, teen, and adult Bryan are three LoRAs, not one.

Layer 3 — pose and reference control at generation time

• ControlNet OpenPose: block the pose (from a posed Blender mannequin or a phone photo of yourself acting it).
• ControlNet Lineart / Canny: paint over a rough drawing and keep its exact composition.
• IP-Adapter: pass the model sheet as an image reference alongside the LoRA to hold costume and colour.
• Fixed seeds + saved settings per shot so any shot can be regenerated exactly.

4. Animation methods, ranked

Honest ranking for this hardware: a single Tesla T4 (16 GB, Turing sm_75, no bf16, fp16 can overflow to NaN in some models), shared with other jobs through a GPU lock.

(a) Cut-out / rigged 2D in Blender — the backbone

Use for: all dialogue, most character acting, and every Looney gag.

• Paint each character in layered parts (head, jaw, eyes, brows, torso, arms, hands, tail) with the character LoRAs, then clean up by hand.
• Rig the parts in Blender 4.5 (Grease Pencil or image-plane puppets with bones). Blender 4.5 is already installed and renders headless on the T4 with EEVEE.
• Rhubarb Lip Sync (free) reads each dialogue line and outputs mouth shapes (A–H, X). A script maps them onto the mouth layer, so lip sync is automatic from the voice track.
• Blinks, breathing, and idle sway are procedural, so every held shot is alive.
• The Looney style wants this: pose-to-pose, smears drawn as extra parts, squash-and-stretch as scale keys.

Cost: mostly your time. Rendering a 5-second cut-out shot is seconds to a minute of GPU (est.). A trained hand can rig a character in a day and animate 30–90 seconds of dialogue a day once the rigs exist (est.).

(b) Image-to-video open models — atmosphere and effects

Use for: fire, water, clouds, the City's light, crowds, Abaddon's drifting ash, the transformation shots. Not for dialogue (mouths drift and faces wander).

Candidates to benchmark in Phase 0, not promises:

Rough expectation: ~3–20 minutes of T4 time per 3–5-second clip at 480p (est.), plus several takes per keeper. Upscale and interpolate afterwards (d). Rule: image-to-video always starts from our approved keyframe, never from text alone, so the style holds.

(c) Depth parallax camera moves — establishing shots

Use for: the City in the sky, the Bronx skyline, Trenton at dawn, Eden, Abaddon's scale shots, and nearly all of v1.

This pipeline is already proven: an SDXL plate → Depth-Anything-V2 depth map → a GPU warp that moves near pixels more than far ones → a real camera move from a still. Cheap (seconds per frame), beautiful, and it never drifts off-model because the painting doesn't change.

(d) Finishing: interpolation and upscale

• RIFE frame interpolation takes cut-out animation on 2s up to smooth 24fps where wanted. Leave Looney shots on 2s, because the choppiness is the style.
• Real-ESRGAN (anime/compact model) upscales 720p renders to 1080p, or 1080p to 4K for a festival master.

Where each shot goes

5. Voice

Scratch track: Piper (free, offline, today)

Piper is installed and runs on the CPU, so the GPU stays free. It's for timing the animatic, not the final film.

Final voices: real people

• Christian voices Christian. It's his life, and nobody else can play it.
• Volunteer actors: friends, family, the Trenton community (with the hospital's and each person's consent, and no real names or likenesses unless they ask for them), local theatre groups, college drama departments. Volunteers get credit, a copy, and a screening.
• Recording kit: one decent USB mic and a closet full of coats. Record to 48 kHz WAV, one line per file, named by script line ID (see §8), so lines drop straight into the edit and into Rhubarb.
• Voice cloning (XTTS-v2, F5-TTS: open, local) is used only with that speaker's explicit, written consent, and only to patch pickups (a line re-worded after the session), never to invent a performance someone didn't give. Nobody's voice is cloned who did not agree, including public figures.

6. Music and sound

The six songs

How: ACE-Step 1.5 runs locally on the T4 (fp32; roughly 5 minutes per 2-minute song, several takes per song). Lyrics are already written in the screenplay. Generate takes, pick the best by ear, align lyrics with whisper, and then animate to the song. Songs are locked before their sequences are animated.

For a commercial release, songs made on Suno must come from a paid plan (the free plan grants no commercial rights). ACE-Step output made locally has no such restriction, but check the model licence before release. Where a real singer will perform, the AI take becomes their guide track.

Score

ACE-Step instrumentals as the temp score in v0/v1. Build leitmotifs early and keep them consistent: - The City: rising, open fifths, choir - Samael / the Ledger: low strings plus a ticking clock, a figure that literally counts - The Heads: kazoo, slide whistle, xylophone (Looney orchestration) - Mamá: guitar, bolero feel - Christian's theme: starts on a lonely 808, ends orchestral

Sound effects

freesound.org CC0 (no attribution needed, free for any use) plus Blender/foley recorded at home. Keep a sfx/LICENSES.csv with the source URL and licence of every file.

7. Shot budget and compute math

Shot count (est.)

~120 pages after the animatic trim × ~11 shots/page ≈ 1,300–1,500 shots, average ~5 s. Animation cuts faster in gags and slower in the Realm, which roughly cancels out.

GPU time per method (all est., to be replaced by Phase 0 measurements)

What that means on the calendar

The T4 is shared. If this film gets ~6 hours of GPU a night on average (overnight, when other jobs are idle), that's ~40 GPU-hours a week.

The GPU is not the bottleneck past v1. People are. Rigging and animating ~700 dialogue shots is the real work, which is why §5 and §10 are about bringing people in.

Free overflow compute

• Kaggle offers free GPU notebooks (typically ~30 hours/week on P100 or 2×T4; check current quotas). That's an extra ~120 GPU-hours/month for embarrassingly parallel batches: LoRA training, image-to-video takes, upscaling. Caveat from experience: recent PyTorch builds may not support the P100's architecture, so pin a compatible build or select the T4 accelerator.
• Colab free tier: occasional T4 sessions, good for experiments.
• Never put private material (unreleased voices, family photos) on a shared notebook without thinking about it first.

8. Pipeline and file layout

Naming

SB_S14_SH030_v003 - SB = Scarlet Beast - S14 = sequence 14 (matches # 14. in the screenplay) - SH030 = shot 30 (numbered in 10s so new shots can go in between: SH035) - v003 = version. Never overwrite; always increment.

Voice lines: SB_S14_L027_christian_take02.wav (L = line number in the sequence).

Folders

scarlet-beast/
  script/            the .fountain files (source of truth)
  shotlist/          shots.csv, generated from the script, hand-edited
  bible/             model sheets, palettes, LUTs, style frames
  lora/              datasets/<character>/, models/<character>_v###.safetensors
  audio/
    voice/scratch/   Piper lines
    voice/final/     real actors
    songs/           takes, stems, chosen masters
    score/  sfx/     + LICENSES.csv
  seq/
    S14/
      SH030/
        plate/       stills, depth maps
        parts/       cut-out layers
        blend/       Blender scene
        render/      v001/, v002/ … (frames)
        SB_S14_SH030_v003.mp4
  edit/              Resolve/Kdenlive projects, EDLs, conform lists
  masters/           exported films: v0, v1, v2 … with date

The shot list is generated from the script

A small script parses the Fountain files and writes shotlist/shots.csv: one row per scene to start (sequence, heading, EARTH/REALM, characters present, dialogue line IDs, songs), then broken into shots by hand in the animatic pass. Extra columns track method (a/b/c), current version, status (v0/v1/v2/v3), duration, seed and prompt.

That CSV drives everything: batch still generation, the Piper scratch track (every dialogue line rendered and named automatically), and an EDL that lays the animatic out in the editor. When a shot is upgraded, the edit picks up the new file because the name pattern is stable.

Editing

On the Windows workstation, DaVinci Resolve (free) is the primary editor and colour tool (the Earth/Realm LUTs live here). Kdenlive is the fallback, and ffmpeg handles batch conforms on the server. Proxies at 720p, finishing at 1080p.

Versioning and backup — take this seriously

The server has known failing drives. Plan as if it will die tomorrow.

• script/, shotlist/, bible/, the rig files, and edit projects go in git (small files), pushed to a private remote.
• Big media is mirrored nightly to the Windows workstation (robocopy/rsync), and monthly to an offline external drive kept unplugged.
• 3-2-1 rule: three copies, two different media, one offline. The masters of every locked song and every finished shot are never in only one place.
• Generated images are reproducible from seed + prompt + LoRA version, so those three are backed up too, not just the pixels.

9. Phases

Risks

10. Distribution: modest and on your terms

1. The site stays unlisted. movie.scarletbeast.com is not linked from anywhere and asks search engines not to index it, until you decide otherwise.
2. Private screenings first. The v0 animatic and v1 motion comic go to a private link for a handful of trusted people. Their reactions decide the v2 priorities.
3. Community screening. When v1 or v2 is ready, a screening for the people it's about and for, the Trenton community among them, if the hospital and the people involved want that.
4. Festivals. Many festivals accept independent and animated features, and some are specifically for outsider, faith, or mental-health storytelling. Expect submission fees; waivers exist. Target after v2.
5. YouTube / free release. The full film free online, clearly labelled, with the songs released alongside, only once you say so.

A finished, honest, strange, beautiful film seen by a thousand people who needed it beats an unfinished masterpiece seen by no one.

Appendix A — The toolbox (all free)

Hardware: one Tesla T4 16 GB (shared), a multi-core Xeon server, ~330 GB free working space on the media volume, and a Windows workstation for editing. Overflow: free Kaggle/Colab GPU hours.

Appendix B — First ten tasks

1. Lock the screenplay draft (read it aloud once; time it).
2. Generate shots.csv from the Fountain files.
3. Render the Piper scratch track for every dialogue line.
4. Style frames: 3 Earth, 3 Realm, 2 Looney.
5. Model sheets: Christian (adult), the Scarlet Woman, Samael, the Beast, Hiss & Growl.
6. Benchmark: one LoRA training run; one image-to-video clip per candidate model; one cut-out test shot with Rhubarb.
7. Stills for sequences 1–5, cut into the animatic.
8. First song takes: "Seven Heads" and "The Ledger".
9. Set up the 3-2-1 backup.
10. First private screening of Act One, however rough.

Come as you are. Leave as you are becoming.