ComfyUI Node
Inkvec Denoise (ConvNeXt)
Restorer for JPEG, WebP and AI-decoder damage on logos and flat artwork (ConvNeXt U-Net, Logolabs/inkvec-denoiser-001), run like `inkvec --restore`. Weights download on first use. auto needs the inkvec binary and a rendered trace (resvg-py).
Inkvec Denoise (ConvNeXt)
- image
- mask
- image
- mask
◄modeauto►
◄deviceauto►
◄tile0►
◄weights_path►
CategoryInkvec
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | — | |
| mode | COMBO | auto | on = always restore. auto = trace once, and restore only if the input disagrees with its own trace where the trace is flat (interior residual above 0.5, the threshold of `inkvec --restore auto`); otherwise the image passes through unchanged. On clean input the restorer costs a few percent of colour accuracy. |
| device | COMBO | auto | auto = CUDA when available, else CPU. |
| tile | INT | 00–4096 | 0 = whole image at once, as the Inkvec engine runs it. 128 = the tiled inference (32 px overlap) the model card's accuracy numbers were measured with; it also bounds memory on large images. |
| maskopt | MASK | Transparency, 1 = transparent (LoadImage's MASK). Passed through. | |
| weights_pathopt | STRING | Local restorer.onnx (or its folder) instead of the download. Empty = download from Hugging Face into models/inkvec/. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| image | IMAGE | — |
| mask | MASK | — |