Video to Article & Timecoded Chapters PHP Script — Free Download (VidPulse AI)

VidPulse AI turns any video or audio file into publish-ready content: interactive timecoded chapters, a full SEO blog article, an X/Twitter thread, a LinkedIn post, a YouTube description and key takeaways — all from a single upload. It is a self-hosted PHP + MySQL app with zero Composer dependencies, built to showcase what Google Gemini's native multimodal engine can do that traditional pipelines can't.

The architectural bet is the interesting part: there is no FFmpeg and no Whisper anywhere in the stack. Instead of extracting audio, transcribing it, and summarizing the text, VidPulse streams the raw MP4/WebM/MOV/MP3/WAV file straight to the Gemini File API using the resumable upload protocol, polls until Google marks the media ACTIVE, and then asks gemini-2.5-flash (or pro) to watch the frames and listen to the audio in one pass. Because the request pins a strict responseSchema with responseMimeType: application/json, the answer comes back as guaranteed-valid structured JSON — no regex parsing, no "the model added prose around my JSON" failures.

The front end is a dark-mode dashboard (Tailwind CDN + vanilla JS): drag-and-drop upload with live status, an HTML5 player where clicking any AI-generated chapter timestamp seeks the video to that exact second, tabbed views for the SEO article and the social media kit, one-click Markdown copy and structured-JSON export, and a MySQL-backed history sidebar so every analysis is stored and reloadable. A one-click web installer checks server dependencies (PHP 8+, PDO, cURL, fileinfo), provisions both tables, and optionally saves your API key.

Everything runs on commodity shared hosting: no root access, no binaries to install, no queue workers. If the host has PHP 8 and MySQL, it runs.

Feature	VidPulse AI (Gemini native)	FFmpeg + Whisper + LLM stack
Server binaries required	None — pure PHP cURL	FFmpeg install (root access), often a GPU box for Whisper
Understands visuals (slides, demos, on-screen text)	✓ — chapters reference what is shown, not just said	✗ — transcript-only, everything visual is lost
Works on shared hosting (cPanel/Plesk)	✓	Rarely — no root, no long-running workers
Output reliability	Schema-enforced JSON (responseSchema) — parses every time	Prompt-and-pray JSON, regex cleanup common
Pipeline steps to maintain	1 API (upload → poll → generate)	3 systems + glue code + temp file management
Silent-video support	✓ — visual-only analysis still works	✗ — empty transcript, empty article
Privacy of processed files	Remote copy auto-deleted from Gemini after analysis; local file stays on your server	Depends on each service in the chain

1. Extract the ZIP into your web directory (e.g. public_html/vidpulse) and edit the database credentials at the top of config.php (DB_HOST, DB_NAME, DB_USER, DB_PASS).

2. Open /install.php in your browser. The setup wizard runs a live server dependency check (PHP version, PDO, cURL, fileinfo, writable uploads/) and creates both MySQL tables with one click. You can paste your Gemini API key here or later.

3. Open the dashboard, drop in a video, pick the output language and model (gemini-2.5-flash for speed, gemini-2.5-pro for depth), and hit Analyze. Chapters, the SEO article and the social kit appear in seconds; every run is saved to the MySQL history sidebar.

Tip: delete install.php after setup, and keep the app behind your own access control if you deploy it publicly — it ships as a single-user tool with no login system.

Heads-up on model IDs: Google retires Gemini model names over time — while seeding the live demo (late Sep 2026) gemini-2.5-flash returned "no longer available to new users" for a fresh API key, and gemini-3.8-flash worked. The model field in Settings is free text, so just type the current model ID if the default ever 404s.

Community build: this one was built by the site's owner with Google Gemini 3.1 Pro, the same way the Claude builds on this page are made — one real brief, the finished code, and the actual prompt published verbatim. Gemini produced the complete application in a single shot (~4 minutes) plus one refinement pass.

The exact prompt given to Gemini:

Build a complete, zero-dependency PHP 8+ and MySQL web application called "VidPulse AI" that showcases Google Gemini's native multimodal video understanding via the Gemini File API (Resumable Upload protocol) and Structured Outputs (responseSchema).

Requirements:
1. Backend (PHP + PDO MySQL):
- Implement a GeminiVideoService class using native PHP cURL that performs a 3-step pipeline:
a) Resumable binary upload of MP4/WebM/MOV/MP3 files to https://generativelanguage.googleapis.com/upload/v1beta/files
b) State polling until the uploaded media status becomes ACTIVE
c) A generateContent call to gemini-2.5-flash with responseMimeType: "application/json" and a strict responseSchema returning: video title, executive summary, clickable timecoded chapters (MM:SS and total seconds), a full SEO blog article (headline, meta description, markdown body, keywords), social media posts (X/Twitter thread, LinkedIn post, YouTube description), and key action items.
d) Automatic cleanup (DELETE) of the remote file on Gemini servers after processing.
2. Database (MySQL):
- Provide database.sql and a one-click web installer (install.php) with a server dependency checker.
- Store API settings in a settings table and all past video analyses in a video_analyses table.
3. Frontend UI:
- Modern dark-mode dashboard using Tailwind CSS CDN and Vanilla JS.
- Drag-and-drop video uploader, interactive HTML5 video player where clicking any generated chapter timestamp jumps the video player directly to that exact second, tabbed views for SEO blog & social media kits, one-click Markdown/JSON export, and MySQL history sidebar.

Independent verification before publishing (done on this site's server, same bar as every other build): all 9 files read line-by-line and linted clean; the resumable-upload / poll / structured-output pipeline in GeminiVideoService.php matches Google's File API contract; fresh install tested end-to-end — dependency checker reported 6/6 PASS, both tables provisioned, dashboard rendered, and an analyze request without an API key failed with a clean, human error message instead of a stack trace. Code quality held up under review: PDO prepared statements throughout, finfo MIME validation on uploads, sanitized server-side filenames, and HTML escaping on both the PHP and JS sides. Real line count is 1,165 across 9 files.

Live demo seeding found one real-world gotcha: with a freshly created API key, Google returned 404 "gemini-2.5-flash is no longer available to new users" — the sample analysis on the demo was generated with gemini-3.8-flash instead (12.5 s for a 54-second video, and its five chapters landed exactly on the five scene changes). The app survives this class of breakage because the model field is editable at runtime.