Qwen Support

[Acly]

Qwen Image

Last updated: 1.44.0

Features

• Qwen Image
• Qwen Image Edit
• Qwen Image Edit 2509
• Qwen Image Layered
• SVDQ Nunchaku
• SVDQ Nunchaku with LoRA
• GGUF
• Inpaint model (but Edit models can already do it)
• Control Layers
• (Multiple) Refernence Images for Qwen Edit
• Managed Install
• Cloud

Requirements

The managed install can run Qwen models since plugin version 1.40.0, but it does not offer automatic model downloads.

For custom ComfyUI installs remember to update:

• ComfyUI
• comfyui-tooling-nodes
• ComfyUI-Nunchaku (optional)
• ComfyUI-GGUF (optional)

Model Download

There are 4 main models:

• Qwen Image - for regular image generation
• Qwen Image Edit - for instruction-based editing
• Qwen Image Edit 2509 - more recent version for instruction-based editing with better support for multiple images
• Qwen Image Layered - for decomposing existing images into transparent layers

Diffusion Models

You can download one or more of these. They go into the /models/diffusion_models folder.

• qwen_image_fp8_e4m3fn.safetensors - Qwen Image, fp8
• qwen_image_edit_fp8_e4m3fn.safetensors - Qwen Image Edit, fp8
• svdq-int4_r32-qwen-image.safetensors - Qwen Image, int4, NVIDIA only
• svdq-int4_r32-qwen-image-edit-2509 - Qwen Image Edit 2509, int4, NVIDIA only
• svdq-fp4_r32-qwen-image.safetensors - Qwen Image, fp4, NVIDIA RTX5000 series only
• svdq-fp4_r32-qwen-image-edit-2509.safetensors - Qwen Image Edit 2509, fp4, NVIDIA RTX5000 series only
• see the Qwen Image Layered section below for additional models

VAE

This is required for all diffusion models.

• qwen_image_vae.safetensors → /models/vae

Text Encoder

This is required for all diffusion models.

• qwen_2.5_vl_7b_fp8_scaled.safetensors → /models/text_encoders

Lightning LoRA

For fp8 or GGUF versions you can add the Lightning LoRA and reduce steps to 8 or 4.

For SVDQ/Nunchaku models LoRA don't work yet, but you can use a merged low-step Lightning version from here:
Qwen Image | Qwen Image Edit | Qwen Image Edit 2509.

You have to create a custom Style for the Lighting versions.

Qwen Image Layered

Qwen Image Layered is a model which takes an existing image and splits it into several layers, separating background, foreground, individual objects and text. These parts can then be edited invidivually with traditional Krita tools.

Diffusion Models

To use it you need the diffusion model (several options with different size below), and the VAE.

• qwen_image_layered_bf16.safetensors - BF16, 40 GB → /models/diffusion_models
• qwen_image_layered_fp8mixed.safetensors - FP8, 20 GB → /models/diffusion_models
• GGUF quantizations - various sizes → /models/diffusion_models

VAE

The VAE for Qwen-Image-Layered is different from the one used for the other Qwen models:

• qwen_image_layered_vae.safetensors → /models/vae

How to use

1. Create a new Style and select the Qwen-Image-Layered diffusion model you downloaded
2. Make sure your canvas is showing the image you want to decompose (not a preview!)
3. Select the number of layers you want as output. You can set it where you'd normally find the strength slider.
4. Click "Edit" to generate the layers. It will take a long time!

Additional information

You can also follow ComfyUI documentation:

• https://docs.comfy.org/tutorials/image/qwen/qwen-image
• https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit

[WatermelonMango451]

Great release.
Regarding "Inpaint model (but Edit models can already do it)," the main reason to add an inpainting ControlNet for Qwen Image is that most LoRAs trained for Qwen Image don't work well with the Edit model. Also, training Qwen Image LoRAs has become much easier since AI Toolkit by Osiris made training possible with around 4 GB of VRAM, and there is no equivalent option for lowering VRAM requirements when training LoRAs on the Edit model.

https://xcancel.com/ostrisai/status/1975642220960072047#m

[StableLlama]

Can you please give an instruction how to use the edit feature with referencing to multiple inputs?
What is "image 1" and "image 2"?

E.g. let's assume the task of placing a person (image 1) and an object (image 2) in a scene / Background (image 3). How should I organize my layers so that this will work?

[StableLlama]

Thank you!
The point that I was missing, is that the layers need to be renamed to "1", "2" and "3" to work like that.

[Acly]

the layer name doesn't matter

afaik image inputs to the text/vision encoder are hardcoded as "Picture 1", "Picture 2" etc in ComfyUI

[StableLlama]

It clearly made a difference for me. I can test again with fixed seed to rule out back luck though.

Trying again with fixed seed I can confirm that the layer names do not have any effect.

[StableLlama]

Based on that I have a feature request to increase the usability here (but I guess it's impossible to implement in a generic way, so I won't be disappointed when it gets declined):
Please add a label that shows the "name" that must be used to reference that image in the prompt (i.e. "Image 1", "Image 2", ...).

Especially in the 3 image situation with two references it becomes guesswork whether the upper one is "Image 2" or "Image 3"

[lokpro]

Canvas = image 1
Reference 1 (top) = image 2
Reference 2 (second) = image 3

that's it.

Reminder: If a referenced layer's image is larger than the canvas, it will be cropped. You need to scale it manually to fit the canvas.

(I’m not an expert, just spent a lot of time recently setting up a ComfyUI myself for the Qwen-image-edit-plus.)

[TijerinArtStudio]

Thank you so much for sharing this post and for all the work you've done to make this a reality. You're doing fantastic work. I'm sure many, many people will appreciate it, even if they don't comment later.

[minecraftman111]

Is there a reason that qwen_image_edit_2509_fp8_e4m3fn.safetensors isn't recognized the same way the other two fp8 models are?

[lokpro]

2509 version is a newer (architecture)

[minecraftman111]

Is there a reason that qwen_image_edit_2509_fp8_e4m3fn.safetensors isn't recognized the same way the other two fp8 models are?

It turned out that this was a unique problem for me due to a poorly upgraded/outdated ComfyUI installation, fixed by reinstalling from scratch, and had nothing to do with "newer architecture".

[Jonseed]

Why does it load the flux text encoders (clip_l and t5)? Shouldn't it use the single clip loader with qwen_2.5_vl_7b_fp8_scaled in qwen_image mode?

[Jonseed]

Btw, if I try just the qwen-image gguf model (not the edit one, but just the regular qwen as above) it also loads the dual clip loader with clip_l and t5.

[Acly]

That indicates it's not a problem with the detection.

The Diffusion Architecture in styles is set to "Automatic" which I'm assuming is what it should be set to since the only other options are "Flux" and "Flux Kontext."

But this sounds like it IS a problem with detection (Flux should not even be options).

I'm not sure. Any chance for some kind of mixup with multiple ComfyUI?
I downloaded the GGUF and it works fine for me.

[Jonseed]

When editing the copy of the Qwen-Edit style, after selecting the Qwen-Edit GGUF models, under Diffusion Architecture I have 3 options showing: Automatic, Flux, and Flux Kontext. It shows the same for regular Qwen too, and for Qwen-Edit 2509 models also, all of which are GGUF. If I select a SVDQ/Nunchaku Qwen model, then it shows Automatic, Qwen, Qwen Edit, and Qwen Edit Plus (so it seems to get those right). So it doesn't seem to be getting the diffusion architecture right. Perhaps it is something with the GGUF detection? I've updated my ComfyUI-GGUF, to be sure, and it still doesn't fix it.

I only have one ComfyUI installed.

[Jonseed]

ok, I just tried clicking the button "look for new checkpoint files," and now the Diffusion Architecture is showing the Qwen options. So maybe it had to rescan the models to get the architecture right?

[Jonseed]

yup, that was it! Now if I dump the workflow, it shows the correct clip loader with the qwen text encoder. So it just had to rescan the model files, it seems, to pick up the correct architecture, but only for the GGUF files. The SVDQ/Nunchaku models worked before I rescanned.

[hansnolte]

For SVDQ/Nunchaku models LoRA don't work yet,...

Hi Acly,

I think, these are the new Loras (without guarantee).
https://huggingface.co/lightx2v/Qwen-Image-Lightning/tree/main/Qwen-Image-Edit-2509

I got the link from here.
https://huggingface.co/nunchaku-tech/nunchaku-qwen-image-edit-2509/discussions/11

Sorry for that question but,
when you write "For SVDQ/Nunchaku models LoRA don't work yet" do you mean the Lightning LoRA or all LoRAs?

I'm currently trying to find out which is the fastest and best quality combination for me (RTX 5600TI 16GB).
There are probably still one or two problems.
#1939 (comment)
nunchaku-tech/nunchaku#728 (comment)

Bye
Hans

[Acly]

You cannot apply any LoRA to SVDQ/Nunchaku models for now.

[StableLlama]

@hansnolte when you want to use the lightning LoRAs you need the pre merged models that were then converted by nunchaku. They are available from them directly:
https://huggingface.co/nunchaku-tech/nunchaku-qwen-image-edit-2509/tree/main

(Please note: rumors say that they had merged the wrong ones. And also there are newer lightning LoRAs available now, that were released about 10 days after the Nunchaku release)

[hansnolte]

You cannot apply any LoRA to SVDQ/Nunchaku models for now.

I was afraid of that, thank you for your answer, Acly.

[hansnolte]

And also there are newer lightning LoRAs available now, that were released about 10 days after the Nunchaku release

Many thanks for the hint StableLlama, but one of my links points exactly to these new Loras.

[StableLlama]

Is there a way to limit the editing to a mask?

I'd like to profit from:

• context control
• far smaller images are sent to the model (my image is 4+ megapixels and thus I run out of VRAM at the moment)

[Acly]

make a selection, as usual

[StableLlama]

I tried that, but it's failing with RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

[StableLlama]

Ah, that seems to be the same as #2114

[Jonseed]

Qwen-Image and Qwen-Image-Edit models are currently not fully compatible with Sage-Attention (particularly version 2.2+ when using the --use-sage-attention flag in ComfyUI), which often results in NaN values and black image outputs during inference. This stems from a buggy Triton backend implementation in Sage-Attention that affects certain GPU architectures (like Ampere sm80 and sm86). To use Qwen models on these GPUs you either have to remove --use-sage-attention from the ComfyUI startup flags, and restart the ComfyUI server, or use a patched version of sage-attention like "Patch Sage Attention KJ" node with sageattn_qk_int8_pv_fp16_cuda option enabled attached to the model loader.

Not sure what the best option is for using Qwen models in Krita. Sage Attention speeds up most models, and only seems incompatible with the Qwen ones. It's very inconvenient to have to remove the sage-attention flag from the startup commands and restart ComfyUI just to use Qwen models. Switching between models quickly becomes impossible. Can the "Patch Sage Attention KJ" node be added as an option in Krita's style settings for Qwen models? KJ Nodes is a pretty popular extension that many users probably already have. If not, users of these GPUs that want to use sage-attention might have to work with custom graphs in Krita where they can apply the patch until these bugs are worked out of Sage-Attention/Triton.

[Jonseed]

Maybe users could just directly patch their sage-attention core.py file for the time being, as in the link above?

    if arch in ("sm80", "sm86"):
        return sageattn_qk_int8_pv_fp16_cuda(q, k, v, tensor_layout=tensor_layout, is_causal=is_causal, sm_scale=sm_scale, return_lse=return_lse, pv_accum_dtype="fp32")
    # elif arch == "sm86":
    #     return sageattn_qk_int8_pv_fp16_triton(q, k, v, tensor_layout=tensor_layout, is_causal=is_causal, sm_scale=sm_scale, return_lse=return_lse)

Mine doesn't look exactly like this. It combines sm75 and sm86 in the same elif statement, but I could add an additional sm86 elif for the cuda line, and remove it from the elif line where it is combined with sm75. But would that mean other models that use sage-attention wouldn't benefit from Triton?

[NoMansPC]

Thank you for all the work you've put into this extension.

Great model. Even with Nunchaku, it takes a long time to output each image, but it's more likely to get it right, so I can keep it running in the background while I do something else.

[fukc-gihtub]

I don't have much success with svdq-fp4_r32-qwen-image-edit-2509. Whatever the prompt is, I get a weird color shift with increased saturation/contrast. (qwen_image_edit_fp8_e4m3fn does the same but a lot slower lol.)

Changing the resolution to 1024x1024 as recommended here on Reddit doesn't help at all.

There are no LORAs or anything. Euler beta, Sampler Steps 20, CFG Scale 2.5

Prompt: The woman is now wearing cat ears. Preserve colors, framing and lighting.

[hansnolte]

Prompt: The woman is now wearing cat ears. Preserve colors, framing and lighting.

Qwen-Image-Edit-2509-Q4_K_M.gguf
LORA - Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors
Euler beta, Sampler Steps 4, CFG Scale 1

cat-ears.mp4

I got this MegaPrompt from here.
https://www.reddit.com/r/comfyui/comments/1ntiher/workflow_qwen_image_edit_2509_outpainting_to_11/

It (usually) ensures that there is no pixel shift or zoom of the generated image.
Works (mostly) with In and Outpainting but apparently not with Remove.

Danamir's "remove the cat" works, but has a Zoom In.
#1939 (comment)

[fukc-gihtub]

Seamlessly extend the scene to fill the padded white bars, outpainting and continuing the existing environment with matching patterns, shapes and textures. Do not modify or repaint the characters’ faces, hairstyles, clothing, or any other distinctive details. Do not change the color and contrast of the image. Focus only on generating new, contextually appropriate background or environmental elements to fill the padded regions. Ensure perfect blending at the edges with no visible seams, artifacts, or style shifts. Maintain the same artistic style, perspective, and overall visual fidelity.

Adding that mega prompt and (most importantly) lowering the CFG Scale to 1.1-1.3 really helps. Thanks.

[agoulddesign]

The 2509 version broke realistic edits from what I can tell. Try original edit model.

Still has too many errors/degradation for how slow it is IMO. Kontext is my go-to for this type of thing

[WatermelonMango451]

InstantX/Qwen-Image-ControlNet-Inpainting is a great inpainting ControlNet for Qwen Image (non-Edit version). It works really well with Lightning LoRAs and other community-made LoRAs, and it complements the Edit model perfectly. The Edit model is great, but sometimes the main model with the inpainting ControlNet gives noticeably better results and more precise control. It uses the same inpainting node as Flux - ControlNetInpaintingAliMamaApply, which is already in Krita AI.

Any chance we might see it added to Krita AI at some point?

Comfy Org Workflow

[Jonseed]

For some modes like "remove content" mode, this would probably work much better than plain Qwen or even Qwen Edit. Currently the inpaint MAT model that the built-in style is using doesn't work well at all.

[Jonseed]

For regular inpainting, this also seems to help blend the change better into the background than plain Qwen.

[Acly]

1.44.0: added Qwen-Image-Layered support, see the special section in top post on how to get it working.

In theorey it could be really useful, but IME it doesn't work that well (and is super slow). But it's very new and might be improved in the future.

[agoulddesign]

"doesn't work that well (and is super slow)"

Not a good mixture. Usually models that don't work very well need a few tries to get close. Probably doesn't get better with strong quant...? Might just hold off until the Nunchaku version

[ThomasEricB]

"Doesn't work that well (and is super slow)"
That might be the understatement of the century! LOL! I'm on a RTX 5090 with Ryzen 9 9950x3d and it's taken well over 15 minutes to render this LOL (still not done!)

[ThomasEricB]

Reduced the image size and the number of layers to 2 since it was taking 30+ to render. Results were deepfried images LOL

[Acly]

For reference, for me 1024x1024 with 4 layers took 10min on RTX 4070 and ~1.5min on RTX 6000 PRO Blackwell

I also got fried image as first of the layers sometimes, usually the other layers were fine though. I think it depends a lot on the content.

[rktvr]

can't get qwen layered to work. latest plugin and updated comfy. using a q6 quant, it always returns a single layer with the original image but deepfried

[ThomasEricB]

Same issue

[Acly]

Just to be sure, the UI should look like this

[rktvr]

aside from the qwen icon being illustrious, it looks the same

[agoulddesign]

New edit model released, apparently addressing some of the issues we have seen.

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

Hopefully they Nunchaku it soon.

[fukc-gihtub]

Key enhancements in Qwen-Image-Edit-2511 include:
mitigate image drift,

[rktvr]

the new qwen edit 2511 requires the FluxKontextMultiReferenceLatentMethod node (comfy core) set to "index_timestep_zero" to not output shifted and oversaturated/garbled images.

[rktvr]

made a feature request/bug report on it: #2245