← back to SERVICE

VC Generator

A local web app that turns a one-line idea into a short K-drama video. AI plans it, draws characters, renders clips, and exports an MP4.

stack
Next.js, React, TypeScript, Tailwind CSS, Codex CLI, gpt-image-2, ComfyUI, MiniMax H3, FFmpeg, Vitest
What I did
  • Built the chat-style workspace and the pipeline from a concept to an exported MP4.
  • Connected the Codex CLI for planning, a local ComfyUI video workflow for clips, and FFmpeg for export.
  • Added Auto and Plan modes, with approve and revise steps in Plan mode, plus stop and resume.

What it is

A single-user web app that runs on one local PC and makes short K-drama videos semi-automatically. The flow:

  1. Enter a concept in one line.
  2. Codex plans the concept, characters, scenes, and video prompts.
  3. $imagegen (gpt-image-2) draws a character sheet for each character.
  4. A local ComfyUI workflow (MiniMax H3) renders each scene clip, using the character sheets as references.
  5. Clips are placed on a built-in timeline, and FFmpeg exports an MP4.

There is no login, database, or cloud queue. All data stays in folders on the PC.

Why I built it

AI video production is split across separate tools: planning, character design, image generation, video prompts, video generation, and editing. Even a short drama means carrying data and context between those steps by hand. This app keeps the whole flow in one place and carries characters and scenes from one step to the next, so their look and identity stay consistent.

How it works

  • One chat screen. Each step’s result appears as a message: concept brief, characters, character sheets, scenes, video prompts, and clips. Every message can expand to show the exact prompt that was sent to the model.
  • Auto and Plan modes. Both run the same steps in the same order. Auto runs from concept to clips with no stops. Plan stops five times for approval (concept, characters, scenes, video prompts, clip generation) and can ask up to three questions at the concept step. In Plan mode, nothing in the next step runs before approval.
  • Revise without starting over. A revised step is saved as a new version, and only the outputs after it are marked as out of date.
  • Uses the existing workflow as is. The app reads the ComfyUI workflow file and never edits it. It only swaps prompts, resolution, length, seed, and reference images in memory. Clip lengths must match the frame counts the workflow can actually produce. Values that don’t match are rejected, not quietly rounded.
  • Stop and resume. Stop interrupts the running job, including ComfyUI. Resume fills in only what is missing: approved steps, finished sheets, and finished clips are kept.
  • Files instead of a database. Each project has a manifest.json written through a temporary file and a rename, so a crash never leaves a half-written file. Backing up means copying the data/ folder.
  • Editable prompts. Creative rules for each step live in prompts/*.md and can be changed without touching code.