Edit Lock

Keeps every part of your picture that an edit did not change exactly where it was, with the colors it had.

An edit model redraws the whole picture, also the parts you did not ask it to change. Those parts often come back a little zoomed in, shifted or tinted. Flip between the original and the edit and the whole picture jumps. Edit Lock stops that: only the edit is new.

It needs no mask. It sees what the edit changes while the picture is being made, and holds on to everything else.

Quick start

Put Edit Lock 🆎 between the model and the sampler: the model (after any LoRA loader) goes in, and its model output goes to the sampler or guider.
Wire the VAE into vae and the picture you are editing into image. Use the same picture the edit model gets, at the size of the result.
After the VAE Decode, add Edit Lock Restore 🆎. Wire the decoded picture into image and Edit Lock's edit_lock output into edit_lock. This puts your own pixels back in the untouched parts.
Wire Restore's report into Show Text 🆎 to read how much was kept.

Run your edit as usual. Prompt, steps and seed stay as they were.

What it does with each kind of edit

A local edit (add glasses, repaint the car, change the sign): the edit happens, and the rest of the picture stays put.
An edit of the whole picture (a new style, new light, fog): there is nothing to keep, so Edit Lock steps aside and the edit runs as if the node were not there. The report says so.
A soft effect over the whole picture (rain, falling snow, haze): it touches everything a little, so there is no untouched part to put back. Edit Lock lets the effect through and only holds the picture in place, so it does not come back zoomed or shifted. The report says so.
A soft effect and a clear change in one edit (rain and an umbrella): the node may lock on the umbrella and hold the rain back everywhere else. If the effect is missing outside the changed part, set Lock to off.
Something in between (a new hair color on a close-up, a new background behind a person): it may lock or step aside. The report says which. These edits often change the light on the face too. auto steps aside for most of them, so the new light comes with the new hair. always keeps the old light on everything it keeps, which can look wrong on a face. A mask settles it: paint the hair and the face, and the rest is locked.
The model ignores your picture and draws a new one: that reads as a whole-picture edit, so auto steps aside. always holds the model to your picture.

Controls

model: the edit model. Patch last, after any LoRAs.
vae: the same VAE the workflow decodes with.
image: the picture you are editing, at the size of the result. If the sizes differ by more than a few percent the lock does nothing, and the report says so.
Lock (lock): auto (the default) locks the untouched parts and steps aside when the whole picture changes. always locks even then. off does nothing, which makes a quick before and after.
Tolerance (tolerance): how big a change must be to count as part of the edit. 1 is the default. Raise it if small unwanted changes get through. Lower it if the lock holds back part of your edit.
mask (optional): white marks where the edit may happen; everything else is locked. With a mask the node does not look for the change itself. A mask from Mask by Touch 🆎 or Mask by Name 🆎 works.

Outputs

model: the model with the lock on.
edit_lock: what the lock kept, for Edit Lock Restore 🆎.

Good to know

The lock belongs to one picture. For an edit with two pictures, give it the one the result should line up with.
It works on single pictures, not video.
The node watches the first steps of a run and locks for the rest, so the model finishes the edit on top of your real picture. Shadows and reflections the edit adds next to the changed part still come through.
If your edit is missing and the report says nearly all of the picture was kept, set Lock to off.
Tested on MiniMax-H3-Image. The lock itself is not tied to one model, but the line between "local edit" and "whole picture" was measured there.

Technical details

At every step an edit model predicts the finished picture. Edit Lock compares that prediction with your picture in latent space.

Watching. The first prediction of a run is a blur, and rain, snow or the light a new hair color throws on a face only show a few steps in. So the node watches until the noise is down to 0.8 of where it began: step 4 of a 6-step MiniMax-H3-Image run. A run that never gets that low decides at its last step. While it watches, the model draws its own picture and only one thing is corrected. The model often draws the whole picture about one percent larger, or a little shifted. The node measures that zoom and shift at every step, on the calmer half of the picture, and moves the prediction back so it sits where your picture does.

Deciding. Change is measured against the quiet part of the picture: the tenth where the prediction differs least. An even tint over the whole picture is taken out first, so a color cast is not read as a change. A spot counts as changed when it differs by 3 times the quiet level, or by 2 times while it touches a spot that does. So a faint change beside a clear one belongs to the edit, and a faint change alone is redraw noise. Tolerance multiplies both numbers.

Locking. From the decision on, wherever the picture did not change, the original latent is put back before the sampler takes its next step. A part once read as changed stays changed for the rest of the run. The last step ends on the original there, so what was kept is the original itself, not a redrawn copy.

Stepping aside. auto does not lock in three cases:

At the first step the whole picture is already changing (a new style, new light). The node reads the calmest fifth of a smoothed copy, where a slightly larger redraw nearly vanishes and a new style does not. On MiniMax-H3-Image, local edits at 1, 4 and 16 megapixels read 0.22 or less there and whole-picture edits 0.27 or more; the line sits at 0.243. The model then runs untouched.
At the decision no place has changed clearly: the strongest spot of the picture is under 4 times the quiet level. That is light rain or fog. Local edits read 4.7 or more there, glasses and makeup included.
At the decision the lock would hand a tenth of the picture or more to the edit, yet the strongest hundredth of the picture is under 4 times the quiet level. That is heavier rain or snow, and a lock would let it through in patches. Local edits that wide read 4.4 or more; rain and snow read 3.6 or less.

In the last two cases the node keeps undoing the zoom and shift to the end of the run, and puts nothing back. An edit close to one of these lines (a hair color change on a close-up, a new background, smoke) can fall on either side. The report says which, and Lock overrides it.

This is the target the TailSteer paper trains a model toward (the source where nothing changed, the model's own edit elsewhere), applied while sampling. It needs no LoRA and no training.

7.2 KiB Raw Permalink Blame History