REC 00:00:00:00 EP.01 — “NEON RUN” · SC 04 · TK 2 4K · 24 FPS · ISO 800

CHAR: VERA · LoRA v3 · IP-Adapter 0.72 Wan 2.2 · dolly-in 0.4

Fully local · RTX 5090 · 32 GB VRAM

Cast once. Shoot forever. Post everywhere.

An automated AI film studio on your own server. You bring the script and the faces. It produces consistent characters, lip-synced dialogue and cinematic edits, then publishes the episode to every channel.

Open the studio console

SCROLL

ComfyUI✦Wan 2.2✦Flux✦Character LoRA✦IP-Adapter✦ControlNet✦LivePortrait✦MuseTalk✦ElevenLabs✦RIFE✦n8n✦Ollama✦Open WebUI✦ ComfyUI✦Wan 2.2✦Flux✦Character LoRA✦IP-Adapter✦ControlNet✦LivePortrait✦MuseTalk✦ElevenLabs✦RIFE✦n8n✦Ollama✦Open WebUI✦

Season 01 · Showreel

Eight episodes. One render server.

Drag the reel to spin it. Click any frame to play it. Every frame is planned, rendered, voiced and cut by the same automated pipeline.

EP.01 — NEON RUN

The pipeline

From script to feed in one run.

n8n orchestrates every stage. Each stage is a versioned ComfyUI workflow called through its API. Failed shots are retried automatically, and nothing gets posted until you approve the cut.

STAGE 01 · SCRIPT

OllamaComfyUITTSFFmpegn8n

Character consistency

One face. Every scene.

Each lead gets a character bible: a trained LoRA, reference set, wardrobe and voice. Every shot is checked against the reference with a face-similarity score, and shots below the threshold are re-rendered automatically.

VERA · 0.94

Vera

Lead · S01

LoRA: vera_v3.safetensors · rank 32
Refs: 24 images · 3 angles · 2 looks
Voice: “Vera” · warm alto · EN / ES
Wardrobe: black leather, red neon key light

Identity match (avg 42 shots)94%

Wardrobe match91%

SC 01 · CAR INTERIOR 0.95

SC 02 · HOOD 0.93

SC 03 · WIDE 0.91

SC 04 · PROFILE 0.96

SC 05 · EXIT 0.92

SC 06 · DOOR 0.94

Dialogue & lip-sync

They speak your lines.

Write the dialogue in the script. Each line gets the character's voice, natural timing and an accurate mouth shape. Captions are generated from the same timeline, so subtitles never drift.

Per-character voicesElevenLabs, or fully local F5-TTS / XTTS clones
Mouth & expressionLatentSync / MuseTalk + LivePortrait retarget
Captions & mixWhisper alignment, music ducking, loudness to −14 LUFS

VERA — line 1400:00.0

Cinematic director tools

Direct the camera, not the prompt.

Camera moves are presets, not prompt luck. Hover a card to preview the move. Every shot then goes through upscaling, frame interpolation and a grade so the series looks like one film.

Dolly inpush 0.4 · 5s

Orbityaw 35° · 6s

Crane uptilt −12° · 6s

Truck leftx −0.3 · 5s

720p RAW4K · UPSCALED + GRADED

24→48Frame interpolation

RIFE / GIMM-VFI for smooth slow motion and clean motion blur.

4KUpscale & detail

SeedVR2 / ESRGAN tiled upscale, with face restore on close-ups only.

LUTOne look per series

Shared grade, grain and letterbox so every episode feels like the same show.

Automated distribution

Rendered at night. Live by morning.

n8n reframes each cut for every platform, writes captions and hashtags with the local LLM, and posts through the official APIs on your schedule. Premium versions run on a separate workflow with their own storage and channels.

ReelsEP.03 drops tonight ✦ #aiseries

TikTokHe knew too much. Part 2 →

48K1.2K3.4K

FacebookThe Archivist — Episode 02

Premium cutseparate workflow · 18+ platforms

TodayChannelAssetStatus

07:30Instagram ReelsEP.03 · 9:16 · 58sPosted

12:00TikTokEP.03 · teaser · 21sUploading 64%

18:00FacebookEP.03 · 16:9 · 2m 40sScheduled

21:00PremiumEP.03 · extended cutAwaiting approval

The machine

Tuned for 32 GB of VRAM.

Ubuntu 26.04, a CUDA 12.8+ build of PyTorch for Blackwell, SageAttention and FP8 weights. ComfyUI, Open WebUI and n8n run as services with a model cache on the NVMe. The plan loads only one heavy model at a time, so nothing swaps mid-render.

RTX 509032 GB · 575 W

Ultra 9 285K24 cores

60 GBsystem RAM

2 TB9100 PRO NVMe

VRAM PLAN · per stageGPU 61°C · 482 W

Wan 2.2 I2V 14B (fp8)~25 GB

Flux dev + LoRA + IP-Adapter~18 GB

LatentSync / MuseTalk~10 GB

Upscale + RIFE (tiled)~9 GB

Ollama 14B (q4) — scripts~11 GB

Figures are estimates for planning and get tuned on the real server.

Delivery plan

Four milestones. One working studio.

Server & stack

Ubuntu, NVIDIA driver, CUDA, PyTorch, ComfyUI, Open WebUI, Ollama, n8n, backups and remote access.

Characters

LoRA training from your references, IP-Adapter/ControlNet workflows, a character bible and a consistency check.

Shots, voice, finish

Video, lip-sync, camera presets, upscale, interpolation, grade, mix and the first full episode.

Automation & posting

End-to-end n8n flows, platform APIs, scheduler, a separate premium path, documentation and handover.

Your first episode, rendered locally.

Open the console to see an episode move through the whole pipeline, from script to scheduled post.