openVideoAI

Local-first toolkit for managing open-source image and video enhancement models.

---

Status: Early development. Model discovery, downloads, and UI configuration are available; media processing and export are not fully implemented yet.

Overview

openVideoAI provides model discovery and downloads, a FastAPI backend, and a browser UI for configuring image and video enhancement workflows. It is in early development: the catalog and UI settings are available, but end-to-end media inference and export are not implemented yet.

Screenshots

Enhancment Page	Models Manager

Features

• Browse enhancement models and manage local downloads.
• Configure enhancement options in the browser UI.
• Explore model metadata and available feature categories through the API.
• Run locally without a cloud dependency.

GPU acceleration

• NVIDIA CUDA — detected automatically; use the CUDA (NVIDIA) backend.

• AMD (Windows) — install torch-directml to enable DirectML acceleration on AMD/Intel cards:

  pip install -r requirements-amd.txt

  Restart the app afterwards. The Settings page will enable the AMD GPU (DirectML) backend and list your adapter. Invalid/legacy device settings never reach PyTorch — an unusable device string falls back to the best available accelerator automatically.

• AMD (Linux) — a ROCm build of PyTorch is detected automatically (hip backend).

• Apple Silicon — Metal/MPS acceleration is enabled on Apple Silicon.

Performance

• Auto-tuned tiling. Real-ESRGAN processes large frames in tiles; each tile is a separate GPU dispatch with fixed overhead (especially on DirectML). When the tile size is left at the shipped default (400 px) the app auto-upgrades it based on your GPU's VRAM (512 / 768 / 1024 px on 4 / 8 / 12+ GB cards), which can cut tiling overhead several-fold. Pick an explicit tile size in Settings to override it.
• Model choice matters. For video, the light 6-block RealESRGAN_x4plus_anime_6B model runs roughly 3–4× faster than the full 23-block RealESRGAN with a modest quality trade-off, and RealESRGAN_x2plus halves the output size when 2× is enough.
• Half precision (Settings → Enable FP16) roughly doubles speed on capable GPUs when the backend supports it.

Model ecosystem

The catalog includes image and video restoration models for upscaling, face recovery, denoising, sharpening, and frame interpolation, alongside color transfer models such as NLUT, Deep Color Transfer, and Reinhard Transfer. It also includes pure-math (non-AI) optical-flow algorithms for frame interpolation (Farnebäck, DIS, Dual TV-L1, Lucas-Kanade, EpicFlow) that run directly on CPU with no model download.

Each entry declares its compute requirements (isAi, requiresCuda, requiresMetal, canRunOnCpu) in models.json. CUDA-required models can only be downloaded on machines with a CUDA-capable GPU, and Metal-required models only on Apple Silicon devices.

See models.md for the complete catalog, comparisons, and usage notes. Catalog entries describe intended integrations; they do not imply that inference is implemented.

Progress

Area	Status
Model catalog and model management	Available
Web UI and API	In progress
Enhancement settings	Available (persisted system, GPU, export, and appearance preferences)
Media inference and export	In progress (upscaling is available; full multi-feature export is planned)
CLI, batch processing, and presets	Planned

For details, see Project status.