FFmpeg Forge 1.1.5 FINAL

✨ New in 1.1.5

🔭 Upscale VisionNext without Visual Enhancer: VisionNext is now built in. Its model is VOSR 2.0 (one step 1.4B generative super resolution, CVPR 2026, Apache 2.0). Click "Install VisionNext model…" once and FFmpeg Forge installs the official VOSR code, the VOSR 2.0 weights with their VAE and DINOv2 encoder, Python and PyTorch CUDA in its own folder (VisionNext next to FFmpegForge.exe), then runs the model itself on your NVIDIA GPU.

🖥️ GPU or CPU: tick GPU (NVIDIA CUDA, fast) or CPU (works without an NVIDIA card, much slower). The choice appears as soon as Upscale VisionNext is ticked.

📈 Upscale factor ×1, ×2, ×3, ×4 and ×8, with tile size 256 or 512. The work grows with the output size: ×3 takes about 2 times longer than ×2, ×4 about 4 times and ×8 about 16 times. Very large outputs switch the VAE to tiles automatically so the GPU memory is not exceeded.

🎬 Videos are encoded directly, never picture by picture. With the built in engine, the video goes from the decoder to the model and straight into the final encoder: no picture files on disk and one single encode. With the Visual Enhancer engine, whose command line only reads pictures, FFmpeg Forge feeds it small batches of frames while the final encoder is already running, so there is still one single encode and no full frame sequence on disk.

⚡ Faster VisionNext: bf16 on the tensor cores with flash attention, all the tiles of a frame processed in one pass, and the same noise for every frame so there is no flicker between frames. Norms, tile blending and the colour fix stay in full precision. Triton is no longer needed (it does not exist on Windows), which fixes the "TritonMissing" error.

🧹 Denoiser IA: the Denoiser section is now a single checkbox. Tick it and FFmpeg Forge removes noise and film grain from videos and images with an AI network, for a clean and natural picture in the spirit of Neat Video. The AI is SCUNet (Swin Conv UNet by Kai Zhang et al., Apache 2.0), trained for real world noise: camera sensor noise, film grain and compression noise. A Neat Video style temporal stabilisation keeps the picture free of flicker.

🚀 Much faster Denoiser IA: the whole SCUNet network is captured once as a CUDA Graph and replayed in a single call per batch, instead of launching about 8,500 small GPU operations one by one. The window attention is fused into one GPU kernel, up to 8 frames are processed per pass, in fp16, with decoding, AI work and encoding running at the same time. If the GPU memory is full, the work size shrinks automatically, and the normal mode is used if CUDA Graphs are not available.

👁️ Live preview: the before / after window closes by itself when the export is finished, and it can no longer make an export fail. If its pictures cannot be written (TEMP folder cleaned by Windows or an antivirus, file locked), only the preview stops and the export goes on.

🛠️ Fixes: clearer error messages for the built in VisionNext engine, a warning when the output is too large for H.264 NVENC (4096 px) or H.265 NVENC (8192 px), and picture sizes that are not a multiple of 8 now work with VisionNext.

✅ Everything from the 1.1.5 betas is included: live before / after preview, DSF and DFF in every channel layout (import and export), SeedVR2, OptiScaler, Scale, NVIDIA RTX AI, CRF NVENC, Spatial Music, 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 very slow.

About 3 GB of disk space for the Denoiser IA, and about 10 GB for VisionNext built in (24 GB of VRAM recommended, more for ×4 and ×8).

×3, ×4 and ×8 need the built in engine. For outputs larger than 4096 px choose libx264, libx265 or H.265.

📦 Downloads

🎒 FFmpegForge-1.1.5-Portable.rar (about 206 MB): everything included, FFmpeg Forge 1.1.5 with FFmpeg 9.0.2 full and ForgeRTX. Extract the whole archive anywhere and run FFmpegForge.exe. The AI Denoiser and VisionNext install next to it.

ℹ️ PyTorch, SCUNet and VOSR 2.0 are not included in the archives: the installers download them from their official sources (GitHub and Hugging Face).