Outpaint workflow that only paints the sides keeps 52.9% of pixels bit-identical

Environmental_Fan600 · reddit · 2026-10-24

Tired of "expand image" workflows repainting the whole photo, the author built a ComfyUI + Qwen Image 2.1 outpaint pipeline that only generates the sides.

Method

Scale the photo to 960px tall, centered on a 2240x960 (21:9) canvas
Pre-fill empty sides with a mirror of the photo edges to prevent a hard grey edge bleeding through the VAE
A soft mask lets Qwen paint the sides plus a 224px ramp over the photo edge for blending
Generation: 25 steps, CFG 1, differential diffusion; the clean photo is used as the reference image (never the padded canvas) because Qwen copies its reference very literally
The untouched original is pasted back at the exact same x/y, fading in over 192px at each edge — the model's version of the photo is thrown away

Pixel preservation (diff vs input)

52.9% of photo pixels bit-identical; the entire center is 100% identical
93.4% within 2/255
A normal Qwen 2.1 Edit "make it wider" pass: only 1.2% identical

Limits & gotchas

Left/right only, no top/bottom yet; you write a one-line description of the new sides
Some seeds leave a visible join — try another seed first; bright objects cut by the frame tend to continue as pale strips, trimming the edge helps
The outer 192px per side is a blend, adjustable via FADE ZONE

Requirements: recent ComfyUI, Qwen Image 2.1 int8, qwen3vl 8b text encoder, 2.1 VAE, ComfyUIessentials, and the author's one-file node pack kirooutpaintstitch. Workflow published on Civitai.

Original post →

Outpaint workflow that only paints the sides keeps 52.9% of pixels bit-identical

More from Multimodal