TrendForge

🎬 TrendForge is an AI-assisted faceless video generator.

Give it a subject, and it can research sources, plan a storyboard, write narration, generate voiceover, create or capture visuals, and assemble a YouTube-ready video.

This project is still evolving. Feedback, experiments, bug reports, and pull requests are welcome.

✨ What It Does

• 🔎 Researches a topic from web/news-style sources.
• 🧠 Separates facts from opinions using an OpenAI-compatible local model endpoint.
• ✍️ Writes structured narration for long-form faceless videos.
• 🗣️ Generates voiceover with Kokoro.
• 🖼️ Creates visuals using source cards, screenshots, and local Stable Diffusion art.
• 👁️ Checks screenshot quality with a local Ollama vision model.
• 🎞️ Assembles the final video with motion, transitions, captions, intro/outro clips, and music support.

🧩 Current Local Model Setup

• 🧠 Research/script/narration model: minimax-m2.5:cloud
• 👁️ Screenshot vision QA model: qwen3.5:4b
• 🎨 Local image generation: Stable Diffusion 1.5 style model, currently tuned for a 4GB GTX 1650 Ti

The vision model is only used to judge screenshots. It does not replace the narration or script model.

🚀 Quick Start

Windows

cd TrendForge
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt
copy .env.example .env

For fully local runs, the stock-media keys can stay as placeholders. Add real keys only if you want Pixabay/Pexels fallback media.

requirements.txt installs the Real-ESRGAN Python dependencies used for higher-quality AI-art upscaling. The model weight file is not stored in Git; TrendForge downloads it automatically on startup when image.upscale_method: realesrgan is enabled.

Then run:

.\run.bat

The UI opens at:

http://127.0.0.1:8510

CLI Mode

.\run.bat --skip-ui --subject "artificial intelligence"

Or directly:

.\.venv\Scripts\python.exe main.py --subject "artificial intelligence"

⚙️ Requirements

• 🐍 Python 3.11+
• 🎞️ FFmpeg
• 🧠 Ollama or another OpenAI-compatible local endpoint
• 🌐 Chrome for screenshot capture
• 🖥️ NVIDIA CUDA GPU recommended for local image generation

The current config is designed to run on modest hardware. Local AI art generates at 768x432, then upscales to video-ready 1920x1080 frames with Real-ESRGAN when available.

🧪 Useful Test Commands

Test Local Image Generation

.\.venv\Scripts\python.exe main.py --image-test "cinematic technology documentary scene" -v

Test Screenshot Vision QA

.\.venv\Scripts\python.exe -m modules.screenshot_vision path\to\screenshot.png --model qwen3.5:4b

🗂️ Project Structure

TrendForge/
  main.py                  CLI pipeline entry point
  server.py                FastAPI web server
  ui.py                    Streamlit UI
  config.yaml              Main app configuration
  requirements.txt         Python dependencies
  models/upscalers/        Auto-downloaded Real-ESRGAN model weights
  frontend/                Browser UI assets
  modules/                 Core pipeline modules
  Assets/                  Branding, intro/outro, docs, optional media
  plans/                   Development notes and roadmap

Generated folders such as temp/, output/, and logs/ are intentionally ignored by Git. Local model files under models/ are also ignored and are recreated/downloaded locally as needed.

🔐 Secrets And Local Files

Do not commit .env.

Use .env.example as the template:

PIXABAY_API_KEY=your_key_here
PEXELS_API_KEY=your_key_here

Local models, virtual environments, logs, generated videos, and temp files are ignored in .gitignore.

🎨 Image Generation Notes

The default local image settings are conservative because this project is being tested on a GTX 1650 Ti with 4GB VRAM:

image:
  width: 768
  height: 432
  upscale_to_output: true
  output_width: 1920
  output_height: 1080
  upscale_method: realesrgan

Native 1920x1080 generation is not recommended on 4GB VRAM. It can pin the GPU and take a very long time per frame.

TrendForge uses SD 1.5 with LCM acceleration by default:

image:
  lcm_steps: 6
  lcm_guidance_scale: 1.8

Real-ESRGAN setup:

• Python packages are installed by pip install -r requirements.txt.
• On first startup or first AI-art upscale, TrendForge downloads RealESRGAN_x4plus.pth to models/upscalers/.
• If the download, dependency import, or model load fails, generation continues with Lanczos upscaling and logs the fallback.
• The model file is about 64 MB and is intentionally not committed to Git.

👁️ Screenshot Quality QA

TrendForge captures source screenshots with Selenium, then applies:

• 🧹 overlay cleanup
• 📄 DOM/source/headline checks
• 🖼️ blank-frame checks
• 👁️ optional Qwen vision scoring

If a screenshot fails quality checks, TrendForge can retry or fall back to a clean branded source card.

🤝 Contributions

Updates are welcome.

Good areas to improve:

• ⚡ Faster local image generation with LCM/SD 1.5 modes
• 👁️ Better screenshot quality scoring
• 🎞️ More polished transitions and motion presets
• 🧪 Automated test coverage
• 📦 Cleaner install and packaging flow
• 🧰 More provider backends for image/TTS/LLM generation

Please avoid committing generated videos, local model files, .env, or virtual environments.

📄 License

MIT