ComfyUI Node
Mosaic Restore
Detect and restore mosaic / pixelated regions in an image or image batch. Multi-frame input is auto-detected and run through BasicVSR++ with true temporal propagation (lada video mode); single frames use pan-augmented test-time augmentation so the flow network has real motion to lock onto. Models are auto-downloaded from HuggingFace on first use.
Mosaic Restore
- image
- image
- mask
◄detection_modelv4-fast►
◄restoration_modelbasicvsrpp-v1.2►
◄deviceauto►
◄fp16true►
◄detect_face_mosaicsfalse►
◄debugfalse►
◄keep_loadedfalse►
Categoryimage/restoration
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| detection_model | COMBO | v4-fast | 4 options: v4-fast, v4-accurate, v3.1-fast, v3.1-accurate |
| restoration_model | COMBO | basicvsrpp-v1.2 | 3 options: basicvsrpp-v1.2, basicvsrpp-v1.1, deepmosaics |
| device | COMBO | auto | 4 options: auto, cuda, mps, cpu |
| fp16 | BOOLEAN | true | Run detection + restoration in half precision. Significantly faster on most GPUs at imperceptible quality cost. Disable on CPU or if you observe NaN outputs. |
| detect_face_mosaics | BOOLEAN | false | If enabled, also restore mosaics that the detector classifies as faces/heads. Off by default to match lada's NSFW-only behaviour. |
| debug | BOOLEAN | false | When enabled the image output is replaced with a 2x2 grid per frame: top-left = source, top-right = restored output, bottom-left = detection regions tinted red, bottom-right = absolute pixel-difference heatmap of what the restoration model changed. |
| keep_loaded | BOOLEAN | false | Keep the detection + restoration models resident in VRAM between runs for faster subsequent invocations. Disable to free GPU memory back to ComfyUI as soon as the node finishes (useful when VRAM is tight or you only run mosaic restoration once per workflow). |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |
| mask | MASK | — |