[NEW] FFmpeg Forge 2.0

✨ FFmpeg Forge 2.0.0

🎨 A brand new design: royal blue and gold premium dark theme, gold logo, rounded cards, gradient buttons and a blue to gold progress bar. The layout stays the one you know from 1.1.5: header on top, your files in the middle and every setting in one column on the right.

🧠 Python / ONNX AI: a new section where you import your own AI models and run them on every frame of your videos and images.
ONNX models (.onnx) run with ONNX Runtime GPU (CUDA) or on the CPU, NCHW or NHWC. Models with a fixed input size are tiled automatically.
PyTorch models (.pth .pt .ckpt .safetensors .bin) are loaded by spandrel: ESRGAN, Real ESRGAN, SwinIR, HAT, DAT, SPAN, SCUNet, NAFNet, Restormer, OmniSR and many more.
Python scripts (.py) with load(device) or process(img). The "Python template" button writes a ready to edit example.
The model scale (×1, ×2, ×4…) is detected automatically. Options: keep the original size, GPU or CPU, fp16 or fp32, tile size. It runs before Scale, RTX, SeedVR2 and VisionNext and keeps your format, codec, audio, trim and filters. "Install AI runtime" installs PyTorch CUDA, ONNX Runtime GPU and spandrel once.

🖼️ PNG/JPG SEQUENCE: in Video settings, tick PNG (lossless, 16 bit when the render is 10 bit) or JPG (quality 2, full 4:4:4 colour) to export your video as pictures. Every frame of the final render is saved as name_000001.png… in a name_PNG (or name_JPG) folder next to the video, with every video option and AI effect applied: Denoiser IA, Python / ONNX AI, Scale, OptiScaler, NVIDIA RTX, SeedVR2 and Upscale VisionNext.

🧹 Denoiser IA models: choose the AI in a list. SCUNet (built in, real noise and film grain), Nomos8k HAT L (finest detail, very heavy), NomosUni DAT (strong denoising, natural look) or PurePhoto SPAN (light clean up, very fast). The ×4 models are brought back to the source size, so the picture is cleaned without changing its resolution.

🚀 Faster Denoiser IA: the whole SCUNet network is replayed as a CUDA Graph (one call per batch instead of about 8,500 small GPU launches), with fused window attention and up to 8 frames per pass.

🔭 Faster Upscale VisionNext: 2 frames per GPU pass on 20 GB cards and above, the bicubic pre resize runs on the CPU in parallel and the colour fix stays on the GPU. The pixels are exactly the same as before. Upscale factor ×1, ×2, ×3, ×4 and ×8, tile size 256 or 512.

🛡️ The live before / after preview can no longer make an export fail: if its pictures cannot be written, only the preview stops and the export goes on.

✅ Everything from 1.1.5 is included: VisionNext built in (GPU or CPU, videos encoded directly), SeedVR2, OptiScaler, Scale, NVIDIA RTX AI, CRF NVENC, DSF and DFF in every channel layout, Dolby Atmos and DTS:X layouts, Spatial Audio, filters and all formats.

⚙️ Requirements for the AI features

An NVIDIA GPU (CUDA) with a recent driver is recommended. Without one, the AI runs on the CPU, which is much slower.

About 3 GB of disk space for the AI runtime (Denoiser IA and Python / ONNX AI), and about 10 GB for VisionNext built in (24 GB of VRAM recommended).

🎒 FFmpegForge-2.0.0-Portable.rar (about 206 MB): everything included, FFmpeg Forge 2.0.0 with FFmpeg 9.0.2 full and ForgeRTX. Extract the whole archive anywhere and run FFmpegForge.exe.

Enjoying FFmpegForge? If you find it useful, please consider stars the project on GitHub! It helps more people discover this free software and supports its continued development.