Qwen Image Edit 2511 - ComfyUI Workflow POWERUP! (Patreon)

Video - https://youtu.be/mCfncH_pZ7I

Overview

Qwen Image Edit 2511 - HuggingFace Page:

https://huggingface.co/Qwen/Qwen-Image-Edit-2511

I hope everyone had a great festive break, and being back from family commitments it is time to get back into even MORE AI image fun!

Qwen Image Edit 2511 is a big improvement over 2509, and I've been having a load of fun with with plain old text-to-image.

It has high-resolution support image right out of the box, beating Z-Image-Turbo's mere 2048x2048 native easily :) You should be able to do 2506x2506 standing, and I even got up to 3k with image and latent upscaling - and that is BEFORE using the 2x VAE (which you can use to crank out 5K images with ease!)

Improved character consistency, better prompt adherence... what's not to like? All available in ComfyUI for the extra POWER and to unleash your creativity :)

Prompting is also very much improved, and there is honestly so much it can do, fitting everything into a single video would take weeks. Face swapping, head swaps, rotations... it all works really well. Do let me know if you find any interesting use cases beyond the norm!

Update 13/Jan/2026 - You may also like this "unblur" LoRA - https://huggingface.co/prithivMLmods/Qwen-Image-Edit-2511-Unblur-Upscale :)

Workflow Information

Three workflows are attached - one with manual masking and one with auto-masking. Personally I prefer the auto mask, but you get a choice :)

There is also one (tti-1024-settings) with the latent image size set to 1024x1024 and none of the options enabled. It is no different to the main workflow at all, it simply has the options set to the most basic mode. The only change is a flip of some switches!

Model Links are available directly in the workflow for you to click on and download:

If you've used Qwen Image Edit before, the only new file is the 2511 model file. The Text Encoder and VAE are the same as with 2509.

Both bf16 and fp8 models are available.

The main NEW things in this workflow are detailed in this note:

The new, main control panel has all the options together in one place, with groups for related components:

Left-click on the switches to turn the named groups ON (yes) or OFF (no).

For example, if you're only generating ONE image (and so you don't want to compare images), set compare image to no:

This will then turn off the group of the same name (Compare Images) OFF. You will see the group change colour.

For contrast, here is the same group turned ON:

As an additional example (one described in the NOTES node), these would be the settings for 1024x1024 text-to-image:

With those settings, no image loading groups are enabled, only the standard KSampler group is ON along with the the standard VAE and an empty latent. Compare is OFF, image upscale is OFF and Latent upscale is OFF.

For AI Enhanced Prompting, the two green groups at the top allow you to pick your Ollama model (see http://ollama.com for downloading ollama & models if you haven't already)

Turn BOTH options on to use the AI prompting. Gemma3:4b is a very small vision capable model, but there are plenty of others to try, such as Qwen3-vl, Ministral and more!

As vision models can "see" images, it is a great way to "re-imagine" existing images without even having 2511 "see" the image video the latent or text encoder nodes.

If you are loading images, then pick how many you'd like to load - 1, 2 or 3:

Images can be encoded using the Qwen Image Text Encoder node, or via the conditioning. The text encoder node will resize everything to 1024x1024, but latent encoding will keep the image resolution.

The empty latent is now also nearby, making it super easy to change if you've enabled it

See the notes node for even more information on the various options:

The new custom KSampler group utilises split sigmas for extra detail and variations

Feel free to use whatever combination of samplers you fancy, though some work better than others! Just don't pick 0 steps for the split sigmas ;)

Whilst you CAN use both sampler groups are once, the output groups are configurable

By default, the Standard VAE output will take the latent from the Standard KSampler, but if you don't have that enabled and are just using the Custom Ksampler, then don't forget to pick the right latent in either the standard vae or 2x vae output group.

As the Latent Upscale Group didn't do too well with masked images, I've also added an IMAGE upscale group

Don't forget you can CHOOSE which image to upscale via the GET node

This means if you want to use the LATENT upscale first and THEN the IMAGE upscale, you are free to do so!

You probably won't want to change anything in the "That be dragons" group at the top of the map, or if you do then it would typically be "set and forget"...

You may wish to change the Ollama options (random seed, temperature, etc) but the defaults are pretty decent.

The negative prompt is also up there, but this won;t have any impact unless you're NOT using the lightning LoRA because that is just how CG1 works!

If you're comparing images, don't forget to select which images you would like to compare using the GET nodes. IMG_1, IMG_2, and IMG_3 are the names for your input images (1, 2 and 3), for example.

I wish everyone A Very Happy New Year, and I wonder what fun 2026 will bring?

Enjoy! :D