Restore mosaic using MiniMax H3 and custom lora.
It needs a beefy GPU. It is slow, and the quality is outstanding. One second of video takes about one minute on an RTX 3070 Ti.
Censored AI-generated clips made with MosaicDiff are on Civitai.
The ready-to-run build is on the Releases page. Download MosaicDiff-windows.7z, open it with 7-Zip, and run MosaicDiff.exe.
That download does not include ComfyUI. Licenses for the shipped pieces are in the licenses folder. Start with licenses\NOTICES.txt.
On the other PC:
- An NVIDIA GPU and a current driver.
- A current ComfyUI install, the normal one from the ComfyUI site or the portable package. MosaicDiff looks for it in the usual folders. If it cannot find it, set ComfyUI Python and ComfyUI folder in Models.
- The Windows download includes the custom LoRA. BasicVSR++, Eros Max, the text encoder, and the video VAE download on first Start when they are missing.
The H3 context-window node and the RTX upscaler ship in comfy_nodes and are loaded from there. The other PC does not install those itself. The first H3 run installs the NVIDIA video effects package into that ComfyUI Python if it is not already there. ComfyUI itself still has to include the built-in MiniMax H3 nodes, so use a current ComfyUI.
The mosaic detector runs on the NVIDIA GPU. The log should say Detector model on CUDAExecutionProvider.
run.bat uses a virtual environment next to this folder when that environment is there. Otherwise it uses Python on the PATH. Finished videos go to Output. If a public weight file is missing, Start downloads it into models\. Place the LoRA there as lora.safetensors. Open Models if a path is wrong.
Install onnxruntime-gpu, not the plain onnxruntime package. The plain package only runs the detector on the CPU, and the two packages conflict. If the CPU package is already installed:
python -m pip uninstall -y onnxruntime
python -m pip install -r requirements.txt
The log should then say Detector model on CUDAExecutionProvider.
Add videos and press Start. Finished files are written to Output. Two sliders matter: how many seconds H3 samples at once, and the H3 resolution. BasicVSR++ runs from its checkpoint, and the mosaic detector runs from its ONNX model. Neither is compiled to TensorRT.