home models images videos 3D Models hubs articles comics crucibles updates shop

Qwen Image 2.1 Edit

Name: Qwen Image 2.1 Edit
Rating: 5 (58 reviews)
Author: AusBoss

1.9k

Updated: Oct 1, 2026

tool

workflow multi reference simple image edit qwen image 2.1

Download

3 variants available

Download Selected

This checkpoint includes a config file, download and place it along side the checkpoint.

Details

Type

Workflows

Stats

952

Reviews

Positive

(35)

Published

Sep 30, 2026

Base Model

Qwen 2.1

Hash

AutoV2

41C03C44D0

About this version

New layout, the same as my other workflows: your image and request on the left, the models beside the steps, before and after on the right. Nodes have their real names.
A LoRA Loader is now the last node in Models, empty. It changes nothing until you add a row. My Consistency LoRA keeps restyles from moving the picture: add it there at strength 1.
The Reference cache node is gone. On its default settings it did nothing, so the picture is the same without it (I compared the pixels). The diffusion model now goes through the LoRA Loader straight to the sampler.
Image Resize is replaced by ComfyUI's own Scale Image to Total Pixels, with the same 1 megapixel, 32-pixel steps and lanczos. ComfyUI counts a megapixel as 1024 × 1024, so the sample now comes out at 768 × 1376 instead of 736 × 1344.
The image loader starts empty, so nothing shows as missing when you open the file. Load your own image.
The note is three steps and four short tips.
Transparent PNGs: the see-through parts are filled with gray before the edit. Before, they came back solid black with a light fringe around the hair.
Same models, sampler settings, seed and starting instruction as v1.0.
Needs AusBoss nodes 2.0.2 or newer.

default creator card background decoration

166

873

AusBoss

Joined Mar 18, 2026

License:

Qwen Research License Agreement

Say what to change in plain words, and Qwen Image 2.1 edits your picture. A slider shows the picture before and after.

Want it with core nodes only? Pick "core nodes only" from the download list (its slider shows in Nodes 2.0). This one doesn't do anything special with my nodes, so it won't hurt my feelings :)

How to use it

Load your image.
Say the change in plain words, like "make it snow" or "change the top to a red knit sweater".
Run.

Don't want custom nodes? There's a core nodes only file in the download list: the same edit with only ComfyUI's own nodes, and the Consistency LoRA ready to switch on. Its before and after slider shows in Nodes 2.0.

What you need

ComfyUI 0.37.0 or newer.
AusBoss nodes 2.0.2 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.
The Qwen Image 2.1 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
The sample picture is in sample_inputs.zip. Put it in ComfyUI/input.

Tips

A style change moves the picture. A painting or anime look can come back zoomed in or shifted, so the slider won't line up. Over 36 restyles I tested, the worst corner typically moved about 24 px. Snow and new clothes stayed within about 10 px.
A LoRA keeps restyles in place. My Consistency LoRA brought that 24 px down to 1.6 px in the same tests. Add it in the LoRA Loader at strength 1. Tested at about 1 megapixel.
Keep the look on a restyle. A painting or anime version can come back in colour or brighter. Say so if you want it kept, like "keep it black and white" or "keep it at night".
Removing a person? Say what they're holding goes too. Otherwise a cup or a phone can stay behind, floating where their hand was.
Raise megapixels for a bigger result. Working image scales your picture to about 1 megapixel first, so a big photo comes back smaller. At 2, my test came out 1088 × 1952 in twice the time.
Don't like the result? Try another seed. Press New seed on the Seed node.
Leave CFG at 1. The negative prompt does nothing at that setting.
About 6 to 12 seconds a run on an RTX 5090. Smaller cards should run slower rather than fail, but I haven't tried one.

Credits

Qwen Image 2.1 is by Alibaba Qwen, under the Qwen Research License: non-commercial use only. Commercial use needs a separate license from Qwen. The files are Comfy-Org's ComfyUI versions.

v1.1 (2026-09-30): new layout, an empty LoRA Loader for the Consistency LoRA, and transparent PNGs no longer come back on black. v1.0 (2026-09-24): first release.

Questions or bugs: GitHub issues or the comments here. More from me on X @Zanzibased and GitHub.