CATEGORY is now "MachinePaintingNodes/Transform", so it renders as
"MachinePaintingNodes / Transform" in the node menu.
"Rotate" is no longer needed in the category, since the display name
carries it - Transform & Rotate Pro is already found by searching
"rotate", and so is the DESCRIPTION text.
One level deep, consistent with Mask, Filter, Color and Blend. README
updated to match.
"Rotate" is the word people actually search for, and almost nothing in
the ecosystem registers under that name, so keeping it in the display
name is worth more than a tidier one. The node menu now shows
"Transform & Rotate Pro", and the class is TransformRotatePro to match.
The registration key stays MP_ImageTransform on purpose. That key is
what saved workflows and placed nodes reference, so changing it again
would break any node already dropped on a canvas. The key is never
shown in the UI, so there is no cost to leaving it as it is - a comment
in __init__.py records why it differs from the class name.
Category is unchanged at MachinePaintingNodes/Rotate & Scale. README
updated throughout.
None of the nodes showed description text in ComfyUI because ComfyUI
reads obj_class.DESCRIPTION when it builds node_info - it does not read
the class docstring, which is what these nodes had.
Every registered node now carries a one-line DESCRIPTION, kept to
29-67 characters so it fits the limited space in ComfyUI Manager and
the node search results. Image Transform Pro's four-sentence
description is shortened to match.
Descriptions are derived from each node's existing docstring, or from
the README for the five classes that had no docstring
(ImageBlendPro, ColorMatchBlend and the three Boolean switch nodes).
README also updated for the recent changes: the New Nodes table now
lists the real "Rotate & Scale" category, the XY placement direction is
documented on both nodes (positive X right, positive Y up), and the
v2.2.0 changelog records the placement convention, the node outputs and
the MP_ImageTransform internal name.
offset_y was following the raw pixel convention where Y grows downward,
so dragging the slider right pushed the image down. Negated it in
build_matrix so both placement sliders now read the same way: drag
right for right/up, drag left for left/down.
One edit covers both nodes, since Image Transform Pro and Remove
Background Pro share build_matrix.
This is the opposite sign to Photoshop's transform Y field, which is a
deliberate choice: on a centre-zero slider "more" reading as "up" is
more intuitive than matching pixel coordinates.
Category is now "MachinePaintingNodes/Rotate & Scale" - a single
subcategory using "&" rather than "/", which ComfyUI would have split
into nested submenus. Keeps the node one level deep like the rest of
the pack. README heading updated to match.
CATEGORY is now "MachinePaintingNodes/Rotate/Scale", so the node sits
under MachinePaintingNodes > Rotate > Scale in the node menu.
Spaces around the slashes are deliberately omitted: ComfyUI splits
CATEGORY on "/" to build the submenus, and a segment of
"MachinePaintingNodes " with a trailing space would not match the
existing "MachinePaintingNodes" root, creating a second top-level menu.
README node-table heading updated to match.
The node does rotate, scale, move, canvas resize, mask cut-out and
background fill, so "rotate" undersold it. "Transform" is the standard
term for rotate + scale + translate, matches the existing
MachinePaintingNodes/Transform category, matches the pack's "... Pro"
convention, and leaves room for flip / skew / corner-pin later.
image_rotate.py -> image_transform.py
class ImageRotate -> ImageTransformPro
MP_ImageRotate -> MP_ImageTransform
display name -> Image Transform Pro
Also adds a DESCRIPTION attribute. ComfyUI reads DESCRIPTION rather
than the docstring for the node tooltip and indexes it for node search,
so the node is still found by typing "rotate", "scale" or "move" even
though those words are no longer in its name.
Free to rename: the node has never shipped in a release, so no saved
workflow refers to the old key.
The mask input was transformed and sent to the mask output, but the
image output ignored it entirely - the image came out as a full
un-masked rectangle, so connecting a mask appeared to do nothing.
The rotated mask now drives what stays visible in the image: whatever
it leaves out is filled with the selected background (transparent,
black or white). The mask is multiplied by the coverage of the
transformed image, so an all-white mask still cannot claim area outside
the image itself. With no mask supplied, behaviour is unchanged - the
transformed image extent becomes the mask.
The in-node transparency-grid preview follows the same visibility, so
what you see matches the image output.
Mirrors the Image Rotate node: a centre-zero, exponential zoom slider
(-100 = 0.25x, -50 = 0.5x, 0 = 1.0x, +50 = 2x, +100 = 4x) sitting
between the angle and the XY placement sliders, and gated by the same
enable_rotation toggle.
The canvas never grows, so zooming in crops at the frame and zooming
out leaves empty edges (image black, mask 0). Image and mask ride the
same matrix, so every output stays registered.
The rotation widgets were hidden by setting `type = "hidden"` and a
computeSize of [0, -4]. That is the older ComfyUI technique and the
current frontend does not support it: there is no "hidden" widget type,
and both isWidgetVisible() and getLayoutWidgets() key off a boolean
`hidden` flag instead.
The result was that the four collapsed widgets still counted as visible,
stayed in the layout pass, and were drawn at zero height on top of the
widgets below them, pushing the node's preview image up over the
sliders. Remove Background Pro on main was unaffected because it has no
hidden widgets.
Hide with `widget.hidden` instead, which drops the widget from the
layout pass entirely. Also swapped the toggle reflow to litegraph's own
expandToFitContent(), which grows the node to clear newly shown widgets
without ever shrinking it, so the preview keeps the room it had.
ImageRotate is a generic internal name that other node packs also
register. ComfyUI merges every pack's NODE_CLASS_MAPPINGS into a single
dict, so identical keys silently overwrite each other and whichever pack
loads last wins - leaving the node missing from the menu.
Renamed the registration key to MP_ImageRotate in both NODE_CLASS_MAPPINGS
and NODE_DISPLAY_NAME_MAPPINGS, matching the existing MP_ImageMaskSwitch
precedent and rule 5 in CLAUDE.md. The Python class name and the
"Image Rotate" display name are unchanged. Safe to rename now because
the node has not shipped in a release, so no saved workflow refers to
the old key.
Image Rotate (new node): rotate, zoom and reposition an image inside a
definable canvas. Free or fixed (45 degree snap) rotation, centre-zero
angle slider (CCW left, CW right), exponential centre-zero zoom
(-100 = 0.25x, +100 = 4x), percent-based XY placement, and
transparent/black/white background fill. An optional mask input rides
the identical matrix so it stays registered to the image. Canvas W/H
inputs default to 0, meaning "match the incoming image".
Remove Background Pro: rotation block added at the bottom of the
settings (enable toggle, free/fixed method, angle slider, XY
placement). Applied after removal so rembg always segments the upright
image; the canvas stays at the original size and image, mask,
masked_image, mask_bw and passthrough are all warped by the same
matrix.
New transform_utils module keeps the rotate/zoom/offset math identical
across both nodes. js/image_transform.js snaps the angle slider to 45
degree steps in fixed mode and hides the rotation widgets until the
toggle is on; Python snaps too, so the result is correct regardless of
frontend version.
New Nodes:
- Mask Composite Pro 2X: combine 2 masks with per-mask blur/grow-shrink/opacity, screen blend, output selector, invert
- Mask Composite Pro 6X: combine up to 6 masks with same features
Updates:
- Inpaint Mask Pro: added expand/shrink, removed SEGS (simplified to image + mask), inpaint_mask now required, multi-color overlay
- Blur Pro: type-specific sliders marked as advanced for cleaner default view
- Widget visibility uses Python "advanced" flag for new ComfyUI frontend compatibility
- Added CLAUDE.md with versioning and registration rules
- Bump to v2.1.1, 40 nodes total
The requirements.txt was missing opencv-python-headless, which is
imported by 7 node files at the top level. When cv2 is not installed,
the entire module fails to load silently, causing KeyError for all
nodes including SeedLock.
Replace fragile top-level `from nodes import EmptyLatentImage` with
a direct torch implementation, fixing plugin load failure. Also update
version string and add _old/ to .gitignore.
MachinePainting