Large edit near the face or a style LoRA at full strength. Add "keep the face identical", lower the LoRA to 0.7, use a consistency LoRA.

Long string or unusual font request. Shorter text in quotes, one line, ask for a plain font; run two or three seeds.

20B model in fp16. fp8 or GGUF Q4 weights and the fp8 text encoder; or hosted.

the best open source ai models

Qwen-Image-Edit guide: instruction prompts, multi-image editing and the LoRAs that make it better

Q: Edit ignored with the lightning LoRA

CFG left at 4 or steps too low for the LoRA version. CFG 1, steps matching the LoRA (4 or 8).

Topic: Prompting

By Captain

Published 2026-05-26

How to get clean edits from Qwen-Image-Edit: writing instructions instead of descriptions, keeping identity and layout, multi-image inputs, text replacement, pose and depth control, the lightning LoRAs for 4-step edits and the anything-to-real and slider LoRAs trending in the Qwen2 family.

Overview

Qwen-Image-Edit is the editing model of Alibaba's

Qwen-Image

line: you give it a picture and an instruction, and it returns the picture with that change and nothing else changed. The 2509 release added several input images (put the jacket from image two on the person in image one), better identity preservation, pose and depth guidance and the same text rendering that made Qwen-Image famous, so you can replace a sign or a label with exact words.

In this catalog the editing generation and its

LoRAs

live in the Qwen2 family. Its LoRAs are among the fastest growing on the site: anything-to-real conversions that turn drawings and renders into photographs, age and body sliders, deblurring, watermark removal, consistency LoRAs that keep a character across edits, and the lightning LoRAs that cut an edit from 40 steps to 4.

The model is 20B parameters, so most people run it quantised (fp8 or GGUF) on 12 to 24 GB or hosted:

PirateDiffusion

exposes it as the edit-qwen workflow (/wf /run:edit-qwen with an input image) and

BitVector

has it in the web app with the LoRAs preinstalled.

Reference

Name	Type	What it is
Instruction prompt	text	An imperative sentence: "Change the shirt to red", "Remove the person on the left", "Replace the sign text with OPEN 24H". Not a scene description.
Image 1 (and 2, 3)	input	The picture to edit; extra images provide a reference object, a style or a second person. Say which image is which in the prompt.
Steps / CFG	setting	40 steps, CFG 4 base. With the lightning LoRA 4 or 8 steps and CFG 1.
Denoise	setting	Keep 1.0; the model decides what to change from the instruction. Lower values blur edits.
Resolution	setting	The model works near 1 megapixel; the ComfyUI template scales the input and back. Very large inputs are downscaled.
Pose / depth control	input	A pose skeleton or depth map as an extra image with "follow the pose of image 2".
Text replacement	prompt	Put the exact text in quotes and name where it goes: Replace the text on the mug with "World's OK Dad".
Lightning LoRA	LoRA	Qwen-Image-Edit-Lightning 4-step or 8-step; weight 1.0; CFG 1.
Style and conversion LoRAs	LoRA	anything2real, anime2real, cinematic, morereal: weight 0.8 to 1.0, added to the instruction with their trigger words.

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Load the ComfyUI Qwen-Image-Edit template (Browse Templates > Image > Qwen Image Edit) and set the loaders: the edit diffusion model (fp8 or GGUF), the Qwen2.5-VL text encoder, the
Qwen
VAE.
Drop the picture into the Load Image node. Write a single instruction. Keep everything you do not mention unmentioned; the model preserves it.
Run at 40 steps / CFG 4 once to see the baseline; then add the lightning LoRA from the Qwen2 family page and switch to 8 steps / CFG 1.
For multi-image edits, add a second Load Image and refer to it: "Put the sunglasses from image 2 on the man in image 1." Reference images work best cropped to the object.
For text: write the exact string in quotes, name the surface, and ask for the same font or colour if you want continuity.
For conversions, add the LoRA and its trigger word to the instruction: "anything2real, turn this illustration into a realistic photograph, keep the composition."
Chain edits: feed the output back in as image 1 for the next instruction instead of asking for three changes at once.
Hosted: on PirateDiffusion reply to a photo with /wf /run:edit-qwen change the shirt to red, or use /workflow /show:edit-qwen to read the fields; on BitVector pick the edit model and upload.

Examples

Single edits

Change the man's jacket to a dark green wool coat, keep everything else the same

Remove the car in the background and extend the hedge

Replace the text on the storefront sign with "CAFE LUNA" in the same style

Two images

Put the handbag from image 2 on the woman's shoulder in image 1, match the lighting of image 1

Conversion LoRA

anything2real, convert this anime illustration into a realistic photograph of the same scene, same pose and framing, natural skin, 50mm lens

PirateDiffusion

/wf /run:edit-qwen /initimage:Ixyz123 change the shirt to red

Tips

One change per instruction gives the cleanest result; stack edits in a chain.
Say what to keep when it matters: "keep her face and hair exactly the same" measurably improves identity.
Lighting edits ("make it golden hour") work; perspective changes ("show it from above") work in 2509 but need a few seeds.
Negative prompts
are weak here; phrase the instruction positively.
For product shots, a plain reference of the product as image 2 beats describing it.
The sliders in the Qwen2 family (age, body, detail) take negative weights for the opposite direction; the model page notes the range.
Watermark and deblur LoRAs are run with the neutral instruction the author gives ("remove the watermark", "deblur this photo") and nothing else.

Name the images in the instruction.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.