Qwen Image Edit - Workflow Power-up! (Patreon)

Video - https://youtu.be/mm2G3hclNCM

Overview

As Qwen Image Edit 2509 is still the best FOSS model, I've powered up the workflow to include some handy new features, including:

Image generation speed up

Image Masking

Automatic Background Removal

Empty Latent option for image re-sizing

Ollama for AI prompting

Speed Up!

If you've got a 40xx series card or newer, the optional Torch Compile node should help speed up image generation after the first one:

Bypass the node (using CTRL+B) or remove it (DEL) if you've got an older card.

CFGNorm can also be bypassed, though it only really seems to make a difference if you are not using a lightning LoRA.

Selection Options

As usual, the various options can be easily enabled or disabled via the RGThree muter (rather than a Bypasser this time!)

Masking

Image 1 has the masking option attached, so if you are going to use a mask then that is the image to apply the mask to!

Using masking is super handy as it means the image doesn't "shift" like it can do without masking :)

Latent Options

You now have 4 options for the latent which gets passed into the Ksampler:

Empty Latent (Use to generate an image of whatever size, up to 2048x2048). Also useful if you are just doing plain text-2-image (no image options selected)

Masked Image (if you've drawn a freehand mask like in the Masking example above)

Image Latent (same as last week's workflow)

Automatic Background Masking (if you're looking to just change the background and can't be bothered to manually set a mask)

AI Prompting

The defaults should work OK, but do take a look at the AI Prompting group if you want to change things such as:

The LLM

The system prompt

Any ollama options, such as seed, temperature, etc

If you use a vision model (such as gemma3) then you can prompt like you would with Qwen Chat. This means you can give instructional prompts such as:

Convert the image style into something else, which must be a mix of three random styles!

Change something in the image into something else

Or, if you're using text to image you can just ask for random things like:

Some type of object from the future - design it yourself

Make a novel hybrid animal

Workflow Information

Still trying to keep the custom nodes to a minimum, but to use the AI Prompting, you will need ollama (https://ollama.com/) installed, and to download at least 1 ollama model.

For example, the vision model I've been using is gemma 3 4b - https://ollama.com/library/gemma3:4b

ollama pull gemma3:4b