Turn a one-line premise into a complete AI manhwa recap video: original story narration → AI Horde images → TTS voice → Ken Burns motion MP4.

Live: https://manhwa-recap-robot.val.run/

File	Purpose
main.ts	HTTP server. Serves index.html/app.js (via staticHTTPServer) and the POST /api/generate AI Horde image backend.
index.html	The app UI.
app.js	All client logic — story writing, strict scripting engine, image/voice/video compilation.

Instead of recapping real manhwa pages, the tool writes an original story from a premise and enforces a hard line-level style:

• Present tense, active voice
• 8–12 words per line
• One action per sentence
• No filler or meta-commentary

How it's enforced in app.js:

• validateLine(line) — heuristic validator that returns { ok, words, issues[] }. It flags violations (past-tense verbs, passive …by…, chained actions, filler phrases) rather than silently rewriting text, since programmatic grammar conversion is unreliable.
• renderValidation(lines) — paints each line in the storyboard green/red with its word count and the exact issue, plus a summary note.
• writeStory — after an LLM call, if more than ~30% of lines fail validation, it retries the LLM up to 2×, feeding back the precise list of failing lines ("Line 2 (14w): 14 words…") so the model corrects itself. One or fewer bad lines is tolerated.
• defaultStory() — the no-key offline fallback. Fixed bug: the original returned a hard-coded 5 lines regardless of the panel count, so narration repeated verbatim. It now builds exactly count distinct lines via combinatorial slot-filling (NARR_OPEN × NARR_WHERE × NARR_THREAT), validated in place, so 12 panels never repeat.

All keyed services are optional (click 🔑 Keys in the app and save → stored in your browser's localStorage):

• LLM Key — any OpenAI-compatible endpoint (sk-… from OpenAI/OpenRouter/Groq) for AI story writing.
• Murf AI Key — paid TTS; leave blank to use the free TTS proxy (tts-proxy).

Server-side env vars (set on the val's Environment Variables page — these are what drive image quality):

Env var	Purpose
GEMINI_API_KEY	(recommended) Google AI Studio key → uses top-tier Gemini Nano Banana (gemini-2.5-flash-image, free-tier friendly; override with GEMINI_IMG_MODEL to try gemini-3.1-flash-image if your plan covers it), with automatic character-consistency (panel 1 is reused as the reference for every later panel).
POLLINATIONS_KEY	Optional. Enables Pollinations flux-anime (manhwa-tuned Flux) + removes its logo.
IMG_ENGINE	Force the engine: gemini | pollinations | horde. Auto-picks by whichever key is set.
GEMINI_IMG_MODEL	Default gemini-3.1-flash-image; override to any Gemini image model.
AI_HORDE_KEY	Last-resort fallback only. Anonymous works but is the low-end path.

main.ts generates one image per panel server-side and returns same-origin data URLs (so the canvas is never tainted), churning a few panels per status poll so each request stays inside the free-tier 1-minute timeout.

Engine order ("no low-end models"): Gemini Nano Banana (primary) → Pollinations flux-anime (secondary) → AI Horde Stable Diffusion (last-resort fallback so the app works with zero keys).

• Gemini generates panel 1 first, then feeds it back as a character-reference image for panels 2–N → the same face, outfit and art style in every panel (real consistency, unlike fixed-seed tricks). Aspect ratio maps from your chosen format (9:16 / 16:9).
• Pollinations caches by URL, so re-polling a finished panel is free; per-panel seeds keep identity stable.
• Horde edge-cases (multiples-of-64 dims, apikey header, 850×850 anonymous cap, model fallback) are preserved on that fallback path only.