Qwen Image 2.1 Edit + Realign

Qwen Image 2.1 edits come back slightly zoomed in or shifted, so they don't sit on top of your picture anymore. This workflow puts every edit back on your picture's frame, with two AusBoss nodes built for the job: a mirrored margin that gives the edit room to move, and Realign to Source, which measures how far it moved and moves it back. Type a short edit and Qwen3-VL writes the full instruction for you, too.

Already have AusBoss nodes? Update them first. This workflow needs AusBoss 2.3.0 or newer. On an older version, the Realign to Source node shows up red as missing. To update: in ComfyUI-Manager, open Custom Nodes Manager, find AusBoss and click Update (or use Update All), then restart ComfyUI.

The first run edits a 912 × 1248 portrait into an 864 × 1184 result in about 12 seconds on an RTX 5090 (about 21 s the first time, while the models load).

How the alignment works

Qwen Image 2.1 redraws an edit with slightly different proportions, by a different amount every seed. Style edits (watercolor, comic, anime) come back about 4% taller on average and sometimes zoomed in 10% or more; lay the edit over your picture and the horizon, a face, the whole scene has moved. Two nodes fix it:

• Realign to Source (new in AusBoss 2.3.0) compares the edit with your picture across the whole frame, works out the zoom (separately across and down) and the shift, and moves the edit back in one step. If the edit only slid by whole pixels, it cuts it out with no resampling at all, so its pixels stay untouched. It finds big zooms too, like Qwen zooming in to fill a padded canvas. On the comic in the gallery it took the edit from 105 px off to 0.6 px; on the watercolor, from 61 px to 0.6 px. Across 166 test edits, style edits went from 35.5 px off to 4.6 px at the median, checked by a separate measuring method.

• A mirrored margin from Load Image + Pad: 64 px of your picture, mirrored outward, on every side. It gives the edit room to move, so nothing drifts off the edge. Without a margin, 13 of 16 test edits came back with an empty strip along an edge after realigning; with the mirror margin, 1 did. Mirror, not gray: flat gray and stretched edge pixels look like a border, and Qwen zoomed in to fill it on 5 and 7 of 16 edits, against 2 of 16 with mirror.

• Together: Realign to Source reads the margin node's stitcher, measures inside your picture's area only (the margin is made up), and cuts the margin off. You get your picture's framing back, with real picture right up to the edges.

• Report says what happened: cut out whole, no resampling means the edit already lined up; warped once plus a zoom means it was moved back; if it names a side and pixels, that edge needed a bigger margin.

Also inside

• Prompt writer: Qwen3-VL 8B, the text encoder Qwen Image 2.1 already loads (no extra download), looks at your picture and your short edit and writes the full instruction: what to change, what it should look like, what to keep. Rewritten instruction shows what it wrote, and Writer instructions is plain text you can change.

• Edit: Qwen Image 2.1 at 25 steps, CFG 1, euler / simple, fixed seed, on a 1 MP working picture.

• Before / After: two sliders, the raw edit against the padded picture and the final result against your picture. Save Image writes a PNG with the workflow inside.

• Consistency LoRA (optional, off): my LoRA that keeps shapes in place on style edits, in a LoRA Loader row you can switch on.

Quick start

1. Update ComfyUI. You need a build with Text Encode Qwen Image 2.1 (I tested on 0.37.0).

2. Install or update AusBoss nodes in ComfyUI-Manager (search "AusBoss", registry id ausboss-nodes) and restart. You need 2.3.0 or newer: that's the version with Realign to Source's margin support and the mirror fill.

3. Drag the workflow in and download the three model files from the card.

4. Load your picture in Source, type your edit in Your edit, and queue.

Models

• qwen_image_2.1_int8_convrot.safetensors → models/diffusion_models · 7.26 GB

• qwen3vl_8b_int8_convrot.safetensors → models/text_encoders · 9.35 GB

• qwen_image_2.1_vae_bf16.safetensors → models/vae · 676 MB

• optional qwen-image-2.1-consistency.safetensors → models/loras · 159 MB

BF16 versions of the Qwen files are on the same page: Comfy-Org/Qwen-Image-2.1. The consistency LoRA's row is off, so the workflow runs without the file.

Custom nodes

Only ComfyUI-AusBoss, 2.3.0 or newer (Manager installs it): free and open source. Text Encode Qwen Image 2.1, the reference cache, Generate Text and String Format are core ComfyUI.

Tips, and what didn't work

• Moving an arm or hand? Say where it ends up. put her arm on the table left her clasped hands in her lap and added a third hand on the table, every time. put her arm on the table, her hands aren't clasped anymore worked. Qwen keeps the old pose unless it's told the old one is gone, and neither prompt writer I tried adds that on its own.

• Check Rewritten instruction when an edit misses. If the writer dropped part of your request, say it more plainly, or wire Your edit straight into the encoder's prompt to skip the writer.

• The consistency LoRA holds poses too. Great for restyles and recolors, where shapes should stay put. Leave it off for pose changes. When it's on, set the margin's four pads to 0.

• I also tried Qwen's own 9B prompt enhancer. It writes richer instructions, but only in thinking mode, which took 20 to 90 seconds per edit, so the workflow uses the Qwen3-VL 8B that's already loaded.

• What Realign can't fix: things the edit redrew in a new spot. It moves the whole picture back, so a restyle lines up closely, not pixel for pixel.

Speed

On an RTX 5090 (32 GB) with the models loaded: about 12 s per edit, rewrite included. The first run takes about 21 s while the models load. The margin adds about 27 % more pixels to render.

I haven't tested smaller cards. ComfyUI offloads what doesn't fit, so expect it to run slower rather than fail.

About the gallery

The PNG is a straight output of this workflow at its saved settings (seed 550181120472374); drag it into ComfyUI to load them. The source picture is my own render. Realign measured this edit within 0.5 px of the original, so it cut the result out without resampling. The cover video shows the watercolor render with and without Realign to Source; the shush video is cut from the gallery PNG's run. The two info cards show one render each, as Qwen drew it and after Realign to Source, measured by the node itself: a comic Qwen zoomed in 9 to 13% (105 px off at the worst corner, 0.6 px after) and a watercolor drawn 5.6% taller (61 px, 0.6 px after). Both pictures are my own held-out renders.

License & credits

Qwen Image 2.1 is © Alibaba Qwen, released under the Qwen Research License: non-commercial use only; commercial use needs a separate license from Qwen. The Qwen3-VL 8B encoder and the model files are Comfy-Org's repackage, under the same license. The consistency LoRA is mine and follows that license when used with the model. The workflow and the AusBoss nodes are free.

Changelog

• v1.0 (2026-09-28): first release.

Questions or bugs: GitHub issues or the comments here. Post what you make with it; I read everything. I post new workflows on X @Zanzibased and GitHub.