Inputs
- Image(s): Input image(s).
AI
AI Nodes Pro
Section titled “AI Nodes ”Image Depth Node Processor
Section titled “Image Depth Node ”Widgets
- Model: Select a depth estimation model from the Model Manager.
Outputs
- Depth Map(s): Grayscale depth map image(s).
Description
Section titled “Description”Estimates the 3D depth of a 2D image using a Depth Anything model. It outputs a normalised grayscale depth map useful for compositing, fog effects, or driving displacement-based nodes.
Parameters
Section titled “Parameters”| Widget | Type | Description |
|---|---|---|
| Model | Dropdown | Depth estimation model to use. Must be downloaded first via the Model Manager before it appears in the dropdown. |
Image Upscaler Node Processor
Section titled “Image Upscaler Node ”Inputs
- Image(s): Input image(s) to upscale.
Widgets
- Model: Select an upscaler model. - Upscale Factor:
2xor4x. - Tile Size: Tile size for VRAM-limited processing.
Outputs
- Image(s): Upscaled image(s). - log: Execution log string.
Description
Section titled “Description”Upscales images using the SPAN (Swift Parameter-free Attention Network) architecture. It supports both 2x and 4x upscaling and includes automatic tiling for processing large images on limited VRAM.
Parameters
Section titled “Parameters”| Widget | Type | Default | Description |
|---|---|---|---|
| Model | Dropdown | None | The SPAN upscaler weights to use. Download via the Model Manager. |
| Upscale Factor | Dropdown | 4x | Output resolution multiplier. Must match the selected model’s trained scale. |
| Tile Size | Int | 0 | Tile size in pixels for tiled inference. 0 = auto (tiles at 512px if input > 1024px). |
Custom Models
Section titled “Custom Models”Any SPAN-compatible .pth file placed in the upscalers/ folder will appear automatically in the Model dropdown. This lets you use third-party or fine-tuned SPAN checkpoints without going through the Model Manager.
Directorymodels/
Directoryupscalers/
- 4x_SPAN.pth
- 2x_SPAN.pth
- your_custom_model.pth
Inpainting Node Processor
Section titled “Inpainting Node ”Inputs
- Image(s): Sequence of frames to inpaint. - Mask(s): Binary mask indicating regions to fill (white = inpaint).
Widgets
- Ref Stride: Reference frame sampling interval. - Neighbor Length: Temporal window size. - Subvideo Length: Chunk size for long sequences.
- Max Resolution: Cap on working resolution.
Outputs
- Image(s): Inpainted output frames.
Description
Section titled “Description”Performs flow-guided video inpainting using FGT (Flow-Guided Transformer). The pipeline uses RAFT to compute optical flow between frames, LAFC (Local-Aware Flow Completion) to fill in flow within masked regions, and a transformer to synthesise the missing content with temporal consistency.
To keep VRAM usage manageable, the node automatically crops to the bounding box of the mask (with a fixed 64 px margin), runs the inpainting at the cropped resolution, then composites the result back onto the full-resolution frame.
Parameters
Section titled “Parameters”| Widget | Type | Default | Range | Description |
|---|---|---|---|---|
| Ref Stride | Int | 10 | 1–30 | Interval at which global reference frames are sampled. Lower values improve quality for long sequences. |
| Neighbor Length | Int | 10 | 2–20 | Number of neighboring frames included in each inpainting window. Higher values improve temporal consistency but increase VRAM usage. |
| Subvideo Length | Int | 80 | 10–200 | Maximum number of frames processed in a single chunk. Reduce if running out of VRAM on long clips. |
| Max Resolution | Int | 720 | 256–1920 | Maximum resolution cap for the working crop. Scales down proportionally if exceeded. |
RIFE Interpolation Node Processor
Section titled “RIFE Interpolation Node ”Inputs
- Image(s): Sequence of frames to interpolate between.
Widgets
- Multiplier: Frame count multiplier (
2xor4x). - Fast Mode (FP16): Enable half-precision for faster inference.
Outputs
- Image(s): Interpolated frame sequence.
Description
Section titled “Description”Increases the frame rate of an image sequence using RIFE v4.6 (Real-Time Intermediate Flow Estimation). It generates synthetic in-between frames by estimating optical flow between each pair of consecutive frames. A 2x multiplier inserts one frame between each pair; 4x inserts three (at timesteps 0.25, 0.5, 0.75). The model is downloaded automatically on first use.
Parameters
Section titled “Parameters”| Widget | Type | Default | Description |
|---|---|---|---|
| Multiplier | Dropdown | 2x | 2x: doubles frame count. 4x: quadruples frame count. |
| Fast Mode (FP16) | Toggle | Off | Uses FP16 (half-precision) on CUDA for faster processing and lower VRAM usage. Has no effect on CPU. |
Video Depth Node Processor
Section titled “Video Depth Node ”Inputs
- Video: Input video.
Widgets
- Model: Select a video depth model from the Model Manager.
Outputs
- Depth Video: Grayscale depth map video. - log: Execution log string.
Description
Section titled “Description”Similar to Image Depth, but processes the entire video using Video Depth Anything. Outputs a grayscale depth map video encoded as H.264 MP4. The output preserves the original frame rate of the input.
Parameters
Section titled “Parameters”| Widget | Type | Description |
|---|---|---|
| Model | Dropdown | Video depth model to use. Must be downloaded first via the Model Manager before it appears in the dropdown. |
Video Upscaler Node Processor
Section titled “Video Upscaler Node ”Inputs
- Video: Input video to upscale.
Widgets
- Model: Select a BasicVSR++ upscaler model.
Outputs
- Video: Upscaled video (4x). - log: Execution log string.
Description
Section titled “Description”Upscales video using BasicVSR++, a recurrent video super-resolution model that leverages temporal information across frames to produce sharper, more consistent results than frame-by-frame upscaling. The output is saved as an H.264 MP4 at CRF 18.
Parameters
Section titled “Parameters”| Widget | Type | Description |
|---|---|---|
| Model | Dropdown | The BasicVSR++ checkpoint to use. reds4 is trained on general video; vimeo90k is trained for compressed video restoration. |