openVideoAI

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

---

Status: Active development. Model discovery, downloads, UI configuration, multi-feature enhancement pipelines, and full export with codec/quality/audio/LUT control are all working.

Overview

openVideoAI provides model discovery and downloads, a FastAPI backend, and a web based UI for configuring image and video enhancement workflows. Multi-feature pipelines (upscale, face recovery, denoise, sharpen, color, interpolation, …) run end-to-end and write video with the codec, profile, bitrate, audio mode, container and LUT you choose in the Export dialog.

Screenshots

Enhancment Page	Models Manager	Export Dialog
		(screenshot pending)

Features

• Browse enhancement models and manage local downloads.
• Build multi-step enhancement pipelines with per-step parameters.
• Export dialog: codec (H.264/H.265/VP9/AV1-NVENC/HEVC-NVENC), profile, Dynamic/Constant bitrate, quality level, container (MP4/MKV/MOV/WebM), audio mode (Copy/AAC/Opus/MP3/None), Apply LUT (.cube), image sequence output.
• Smart Select: auto-pick the best model for the loaded file (media, memory, install state).
• Settings persisted to disk (system resources, GPU backend, export defaults, appearance).
• Native desktop window via pywebview; also runs in a browser tab.

Progress

Area	Status
Model catalog and model management	✅ Available
Web UI and API	✅ Available
Enhancement settings	✅ Persisted (system, GPU, export, appearance, models, advanced)
Media inference and export	✅ Working (all pipelines, real ffmpeg encoder, audio muxing, LUT)
CLI, batch processing, and presets	🔄 Planned

For details, see Project status.

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) that run directly on CPU with no model download. Farnebäck, DIS and Lucas-Kanade work out of the box; Dual TV-L1 additionally needs opencv-contrib-python.

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. Entries may also carry a media field (image, video or both, the default) marking models that apply to only one kind of media; the frame-interpolation and video-encoder entries are video-only. The Smart Select switch in Settings → Models uses those fields, the detected hardware and the memory budget in Settings → System to choose the best model for the file being processed instead of the catalog default — see models.md. Each entry also carries a bestFor note on when to reach for it and which sibling to prefer instead; the model details popup shows it, together with the media label, under Best used for.

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