CNN
DETERMINISTICRCAN, EDSR
Classic convolutional networks. Deterministic, bit-identical output across runs. Best for archival work.
VideoForge
Clean up, upscale, crop, and export videos locally on Windows. Your GPU does the rendering, your footage never leaves the machine.
VideoForge is a local-first desktop app that runs old, soft, or low-resolution video through a nine-stage processing pipeline: repair the source, upscale it with the right AI model for the job, reframe it, grade it, and export it with hardware NVENC encoding. A Rust orchestration layer, a Python inference engine, and an optional GPU-native video engine sit behind one Tauri app.
- Windows 10/11
- NVIDIA CUDA 11.7+
- Local-only
- Deterministic modes
- In development
Processing pipeline
Nine stages, import to export
The app is organized the way the work actually flows. Every stage is a panel in the settings rail: inspect it, enable it, or bypass it, then hit render.
- 1
Import Source
Drop in a local video or image. Resolution, aspect, and format are probed on the spot.
- 2
Media Inspector
Container, codec, frame rate, and stream details, read locally with FFprobe.
- 3
Source RepairBYPASS
Denoise, deblock, deband, deflicker, deinterlace, and chroma noise cleanup before upscaling.
- 4
AI UpscaleCNN
CNN, GAN, Transformer, Diffusion, and Lightweight model families with 2x to 4x scale factors.
- 5
Crop & Frame
Interactive crop overlay on the preview, with black bar detection.
- 6
Transform
Rotation and flip or mirror, applied in the export chain.
- 7
Color GradingAUTO
Brightness, contrast, saturation, and hue, plus an auto-grade analysis flow.
- 8
Final Sharpen
Output sharpening tuned after the upscale, not before it.
- 9
Export Output
Hardware NVENC H.264/H.265 encode straight to MP4 on your disk.
Source repair
Fix the source before you upscale it
Upscaling a noisy, blocky source just gives you sharper noise. VideoForge puts cleanup ahead of inference, each pass with its own level control.
Denoise, deblock, deband, deflicker, deinterlace, chroma noise cleanup, pre-upscale sharpening, and black bar detection, all previewable per frame or per clip range.
AI upscale
Five model families, honestly labeled
Pick the architecture that matches the job. Deterministic models are marked, because for archival work the same input should always produce the same output.
GAN
NON-DETRealESRGAN
Generative detail synthesis for heavier restoration, clearly labeled as non-deterministic.
Transformer
NON-DETSwinIR, HAT, Swin2SR
Attention-based models for the full pipeline plus secondary model blending.
Diffusion
NON-DETDiffusion SR
Diffusion-based enhancement for cases where synthesis quality matters most.
Lightweight
NON-DETFast preview models
Small CNNs for quick previews before committing to a long render.
Engine routes
One app, three execution paths
Jobs route to the engine that fits the model and media. Eligibility is checked up front, and the Python path remains the default when the native family is not the right match.
Python sidecar
DEFAULTThe broadest-compatibility route: FFmpeg decode, shared-memory ring buffer, PyTorch inference, FFmpeg encode. Also carries the research layer: multi-model blending, frequency band analysis, and hallucination detection.
Native direct
FASTESTIn-process engine-v2: NVDEC decode, CUDA preprocessing, TensorRT or ONNX Runtime inference, NVENC encode. Fully GPU-resident with no CPU round-trips.
Native CLI-backed
FALLBACKThe native-family compatibility route through the rave adapter, used when the direct path is not the right fit.
Core philosophy
Local, deterministic, under your control
- Video processing happens locally on your PC and GPU, footage is never uploaded for processing
- Supported models (RCAN, EDSR) produce bit-identical output across runs
- A deterministic precision mode forces cudnn.deterministic and disables TF32
- Full user control over trim, crop, grading, model selection, and precision, with preview before commit
- Optional per-run artifact bundles record exactly what settings produced an output
Under the hood
The stack
- Tauri 2.0
- Rust + Tokio
- React 19
- Zustand
- PyTorch + CUDA
- FFmpeg NVDEC/NVENC
- TensorRT / ONNX Runtime
- Zenoh + shared memory
Frames move between Rust and Python through a zero-copy shared-memory ring buffer with Zenoh signaling. The native engine-v2 path keeps the whole pipeline resident on the GPU.
Requirements
Built for NVIDIA Windows machines
- Windows 10 or 11
- NVIDIA GPU with CUDA 11.7+ and NVENC support
- Model weights load from a local weights folder, scanned automatically at startup
Current status
VideoForge v0.1.0, in development
VideoForge is being built in the open where it can be: the GPU engine work, the run-artifact tooling, and the deterministic execution research are public repos. The app itself is pre-release and not yet downloadable. It will ship as a freemium desktop app: a Free tier for previews and watermarked sample exports, and VideoForge Pro for full-length, full-resolution work.
Old footage in, clean footage out
VideoForge is the deep end of the same idea behind the Fast Series: practical Windows tools that do the heavy lifting on your own hardware. Want to know when it ships, or want early access as a tester?