ComfyUI Node
kubakub onnx style transfer
Restyles an image with your own trained pix2pix model (an .onnx file), at full resolution by working in overlapping tiles. Runs on the GPU when onnxruntime has CUDA, else on the CPU.
kubakub onnx style transfer
- image
- styled_image
◄model_path►
◄modetiled►
◄tile_size512►
◄overlap64►
◄blend1.00►
◄passes1►
◄keep_loadedfalse►
Categorykubakub/lab
Inputs (8)
| Name | Type | Default | Description |
|---|---|---|---|
| image | IMAGE | The image (or batch) to restyle; each image is done on its own. | |
| model_path | STRING | An .onnx file: a full path, or a name inside ComfyUI/models/onnx. | |
| modeopt | COMBO | tiled | tiled runs the model on overlapping squares at full resolution. single squeezes the whole image into one square and scales it back (fast, softer). |
| tile_sizeopt | INT | 512256–1024 | Square size in pixels the model works at. Use the size it was trained at (usually 512). |
| overlapopt | INT | 640–256 | tiled only. Pixels neighbouring tiles share and blend across, so no seams show. |
| blendopt | FLOAT | 1.000–1 | Mix with the original: 1 = fully styled, 0 = the original image. |
| passesopt | INT | 11–5 | Run the model again on its own result this many times for a stronger style. 1 = once. |
| keep_loadedopt | BOOLEAN | false | Keep the ONNX session (and its VRAM) between runs; off frees the GPU for the diffusion models. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| styled_image | IMAGE | The restyled image, same size as the input. |