Remaps image luminance or a mask through draggable RGBA stops, with Linear, Ease, Constant, Cardinal and B-Spline interpolation across RGB, HSV or HSL, and Blender-style hue directions. Source alpha is preserved exactly.

It also forced something larger: a reusable inspector-control registry, so a plugin can be handed a real gradient editor with stop dragging and an app-wide eyedropper without the app ever executing plugin JavaScript. ColorRamp was its first tenant. The Pose Editor was its second.

Depth Anything V2 running locally on the GPU, loaded lazily and downloaded once. Small, base and large weights, with a resolution control for when edge quality matters more than the second it costs.

Free, offline, and accurate enough to composite against rather than merely look at, which is the whole reason it is worth having on the machine instead of behind an API.

Detection returns 133 whole-body points including individual finger joints, rather than the eighteen most pose tools settle for. Coordinates are stored normalised, so a pose captured once stays correct at any output size.

The editor draws that skeleton on a canvas and lets me drag a joint to change it. ControlNet generation then runs against the edited pose, borrowing the already-resident local model instead of loading a second copy.

Then I measured it. Roughly 3.8 seconds per step at full resolution, close to two minutes for one image. That number is what turns a feature into an informed choice about when to reach for it.

Every generative node keeps its own results. Arrow through them, pick the one that works, and it becomes the active output downstream, replayed locally without a second paid request.

History stays out of the saved workflow on purpose. A workflow is a recipe, not a scrapbook. And a duplicated node starts clean rather than inheriting a paid result it never earned.

Some upscalers leave a faint repeating grid behind. This carves the offending frequency out of the image’s own spectrum, with manual period targeting and optional diagonal checkerboard removal, across 8-bit and 16-bit with alpha and ICC preserved.

While generalising it past one fixed spacing, testing turned up a real gap: full cancellation only happened at the single period it was first tuned for, and everything else silently landed at half strength. After the fix, the targeted frequency fell by 99.8 percent against a real photograph, with error down 84 percent. Shipped as one plugin file, zero core edits.

Cuts a figure from its background using pinned BiRefNet Lite salient-object segmentation. Optional painting only selects which predicted component I meant, and never leaks into the output itself.

This replaced an earlier whole-image GrabCut with optional SAM refinement. The simpler pinned model was chosen after checking whether the heavier one actually earned its size. It did not.

A freeform writing field where every connected prompt appears as a coloured block I can drag to any caret position inside my own text. The merged result is what goes downstream.

Getting the caret behaviour right took three attempts. Editing text around inline non-editable blocks is where browsers disagree most, and the fix that finally held came from extracting the logic into a testable module rather than patching the editor in place.

Background work split into a general lane and a serial accelerator lane. Independent CPU and provider work keeps moving while a local GPU model runs, and GPU-heavy inference stays serialised so it never fights itself for memory.

Queued work can be cancelled with a guarantee it will not start afterwards, and running work stops at the next safe boundary rather than mid-write.