Wires between nodes
Many nodes have small ports on their edges: inputs on one side, outputs on the other. Drag from an output to an input to draw a wire, and files flow along it. A Folder node, for example, has an All files output that sends its files on.
A wire that doesn’t make sense is refused, and a label tells you why (“doesn’t accept”). Loops aren’t allowed. Double-click a port to give it a name.
The pipeline nodes
They’re in the node picker, under Pipeline.
- Folder source is where a pipeline starts. Pick a folder with Browse; it sends on every file in that folder, not its subfolders. Refresh reads the folder again.
- Filter keeps the files with the extensions you list, like
jpg, png, webp(no dots; commas, semicolons or spaces between them). - AI Classify looks at each image with the local AI and gives it a category, some tags and a short subject. The categories are landscape, portrait, food, architecture, nature, document, screenshot, art, event, object and other. When the AI doesn’t give one of these, the image gets no category rather than a guess. Press Classify to run it. It needs the AI engine, see Local AI.
- Rename renames the files with a pattern. Click the variables to insert them:
{name},{ext},{idx}(a running number),{date},{category}and{subject}(these two come from AI Classify). An image with no category is nameduncategorized, and characters Windows doesn’t allow in file names are taken out of the subject. - Move moves the files to a folder you choose. It moves, it doesn’t copy.
Preview first
Every node shows what it would do: which files pass, and for Rename and Move the new names and the new paths. Nothing changes on disk until you press Run and confirm. The confirmation says exactly how many files will be renamed or moved, and where.
A small badge on each node tells you its state: running, waiting, out of date, error.
Undo after
Renames and moves made by a pipeline are recorded in History, and you can undo them from there. See Folders and files. If a name was already taken and Windows picked another one, History knows the real name.
Clean up and Upscale
Two nodes for photos, also in the picker under Pipeline. Clean up removes JPEG artifacts (the blocks and the halos along edges) and keeps the photo the same size. Upscale makes it 2x, 4x or 8x bigger. Connect them to an Image node to work on one photo, or to a Folder node to do them all.
Unlike the other pipeline nodes they have no preview. They only run when you press their button, and they leave the original alone because they write a new file.
- In Clean up, Strength is Automatic (the model works out how hard the photo was compressed) or Manual. In Manual the slider goes from light to strong, and the number next to it, from Q 100 down to Q 10, is the JPEG quality the model assumes.
- In Upscale you pick the model. Photo is for real photos and gives natural detail. Fast is light and also removes grain; Noise sets how much (low, medium, high). Faithful keeps shapes and lettering where they are, but it’s slow on large photos. Then pick the Factor. When the photo’s size is known, the size that will come out is shown below.
Chained, Image → Clean up → Upscale, the two nodes work in one pass. The Clean up and upscale button sits on the last one, and you only get the final result, with no cleaned-up file in between. The first node keeps a Clean up only button in case you want to stop there.
The new file takes the original’s name plus a suffix: _clean, _4x (or _2x, _8x), _clean_4x when chained. You can type a different one in Suffix, and reset goes back to the automatic one. If the name is taken Warpstead adds -1, -2; it never overwrites. The file goes next to the original, or into the folder of a Folder node connected to the output, and the bottom of the node tells you which.
The format stays the same: JPEG and WebP at quality 95, PNG and TIFF at their bit depth, with the original’s EXIF, XMP (ratings included) and color profile. BMP, GIF, AVIF, JPEG XL and PSD come out as PNG, and RAW files as 16-bit TIFF with your adjustments applied. HEIC files can’t be opened yet and CMYK photos are skipped; either way the node says why. If a result would go over 200 megapixels, Warpstead asks before it starts.
Everything happens on your PC, and no photo leaves it. The models run on the graphics card; without one, the node warns that it will be slow and uses the processor. Each model downloads by itself the first time it’s needed (2.4 to 144 MB), then the work starts. While it runs the node shows which image it’s on, and Cancel stops everything.
Before and after
As soon as you connect a photo, the node takes its shape and shows it at the top; connected to a Folder it shows the first photo and how many there are. When the run is done the comparison opens right there, with a line you can drag: Before on the left, After on the right. Before is the original brought to the same size the ordinary way, so you see what the model added. Double-click the handle to put it back in the middle, and once the handle is focused the arrow keys move it. From a folder, a strip of the results sits under the photo: click one to compare it. Open in a node, on the node’s “Done” line, opens the result you’re looking at in an Image node. Run it again and the comparison updates.
Other models
Settings → Engines → Clean up and Upscale lists the models, their size and whether they’re downloaded. Remove frees the space, and the model downloads again next time. Add model…, also in the Upscale model menu, lets you pick your own .onnx file: Warpstead copies it, tries it, and puts it under Clean up if it’s 1x or under Upscale if it’s 2x or 4x. Check the model’s license, because many models on OpenModelDB can’t be used for commercial work.
An example
Folder source on your Downloads → Filter jpg, png → AI Classify → Rename {category}-{idx} → Move to Pictures\Sorted. Check the preview on Rename and Move, then Run.