🎬 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 evolving. Feedback, experiments, bug reports, and pull requests are welcome.
- 🔎 Researches a topic from web and news-style sources.
- 🧠 Separates facts from opinions with an OpenAI-compatible model endpoint.
- ✍️ Writes structured narration for long-form faceless videos.
- 🗣️ Generates voiceover with Kokoro.
- 🖼️ Creates visuals from source cards, screenshots, and local SDXL art.
- 👁️ Checks screenshot quality with a local Ollama vision model.
- 🎞️ Assembles motion, transitions, captions, intro/outro clips, and audio into a final video.
The default configuration is tuned for an NVIDIA GPU with 8 GB VRAM and a computer with 16 GB system RAM:
| Task | Default model | Notes |
|---|---|---|
| Research, script, narration | minimax-m2.5:cloud |
OpenAI-compatible Ollama cloud model |
| Screenshot vision QA | qwen3.5:9b |
6.6 GB Q4 multimodal model; unloaded after each check |
| Image generation | Lykon/dreamshaper-xl-lightning |
SDXL Lightning, FP16, 4 steps |
SDXL renders AI-art frames at 1152x896, then TrendForge upscales them 2x to 2304x1792. Model-level CPU offload and VAE tiling reduce peak memory use.
The vision model and image model are not kept in VRAM together. vision_keep_alive: 0 tells Ollama to release Qwen after each screenshot check so SDXL has the GPU available.
- Windows 10 or 11
- Python 3.11 or 3.12
- NVIDIA GPU with CUDA support and 8 GB VRAM recommended
- 16 GB system RAM
- Current NVIDIA driver
- FFmpeg available on
PATH - Ollama
- Chrome for screenshot capture
Clone and enter the repository:
git clone https://github.com/IsaacNewtonne/TrendForge.git
cd TrendForgeCreate the environment and install Python dependencies:
python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip
pip install -r requirements.txt
Copy-Item .env.example .envConfirm that PyTorch can see the GPU:
python -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'CPU only')"If that reports False, install the CUDA build of PyTorch using the command generated at pytorch.org/get-started/locally, then run the check again.
Install the local screenshot vision model:
ollama pull qwen3.5:9bThe research/script default uses minimax-m2.5:cloud. Sign in to Ollama if required by your Ollama installation. To use another OpenAI-compatible model, edit opencode.model and opencode.base_url in config.yaml.
Start TrendForge:
.\run.batOn every launch, the bootstrap checks the main Python requirements, Ollama, FFmpeg, Chrome, and the required Ollama models. Missing components are installed automatically. The first launch can therefore take considerably longer while packages or models download.
The UI opens at http://127.0.0.1:8510.
TrendForge checkpoints each completed pipeline stage under temp/checkpoints/. Running the same
topic again with the same effective configuration restores completed research, analysis, script,
voiceover, and visual stages. Audio and image assets are copied into the run-specific checkpoint so
another topic cannot overwrite them. An interrupted stage is rerun automatically.
To deliberately rebuild every stage from scratch, use CLI mode with --no-resume.
The first AI-art run downloads the FP16 DreamShaper XL Lightning model from Hugging Face into the local model cache. Model files are ignored by Git.
.\run.bat --skip-ui --subject "artificial intelligence"Or run the Python entry point directly:
.\.venv\Scripts\python.exe main.py --subject "artificial intelligence"Test local SDXL image generation:
.\.venv\Scripts\python.exe main.py --image-test "cinematic technology documentary scene" -vTest screenshot vision QA:
.\.venv\Scripts\python.exe -m modules.screenshot_vision path\to\screenshot.png --model qwen3.5:9bRun the test suite:
.\.venv\Scripts\python.exe -m unittest discover -s testsThe main 8 GB profile lives in config.yaml:
image:
engine: sdxl
model_id: Lykon/dreamshaper-xl-1-0
variant: fp16
scheduler: deis_multistep
acceleration: none
dtype: fp16
vae_id: madebyollin/sdxl-vae-fp16-fix
vae_dtype: fp16
steps: 25
guidance_scale: 6.0
width: 1152
height: 896
require_native_resolution: false
upscale_to_output: true
upscale_scale: 2
upscale_method: realesrgan
detail_pass:
enabled: true
steps: 12
strength: 0.22
guidance_scale: 5.0
enable_cpu_offload: true
enable_attention_slicing: false
enable_vae_tiling: trueThe final-quality profile uses full DreamShaper XL with its recommended DEIS scheduler, an FP16-safe SDXL VAE, a restrained img2img detail pass, and Real-ESRGAN upscaling. Model CPU offload and VAE tiling keep the workflow within the RTX 4060 Laptop GPU's 8 GB VRAM budget. Torch 2 already uses memory-efficient scaled-dot-product attention, so additional attention slicing is disabled by default.
If image generation runs out of VRAM, close other GPU-heavy programs first. For a last-resort lower-memory mode, set image.low_vram_mode: sequential_cpu_offload; it is considerably slower than the default model offload.
TrendForge/
main.py CLI pipeline entry point
server.py FastAPI web server
ui.py Streamlit UI
config.yaml Main app and hardware profile
requirements.txt Python dependencies
frontend/ Browser UI assets
modules/ Pipeline modules
tests/ Unit tests
Assets/ Branding, intro/outro, and docs
Generated temp/, output/, and logs/ folders are ignored by Git. Local model files under models/ are also ignored.
Do not commit .env. Copy .env.example and add real Pixabay or Pexels keys only if you want those fallback media providers. Fully local runs can keep the placeholders.
MIT