Editing images often feels like a hectic work and time consuming. Portraits lose facial identity during transformations. Products lose originality when placed into posters or advertisements. Text changes disturb design flow instead of blending naturally. Working with more than one image magnifies the struggle because combinations such as person with person or person with product rarely look consistent. Qwen-Image-Edit-2509 (Plus) can handle these challenges directly. Creators want edits that preserve identity, maintain details, and remain visually seamless.
The model is designed to keep edits consistent across both single-image and multi-image scenarios. It aims to deliver natural blending when combining images while preserving details when making adjustments to a single image.
 |
| Qwen-Image-Edit-2509 showcase (Ref-official page) |
The foundation of Qwen-Image-Edit-2509 is the Qwen-Image-Edit architecture. Multi-image editing becomes possible through training with image concatenation which allows the model to understand relationships between multiple inputs. Performance is at its best with one to three images. Single-image consistency has been refined in three specific areas.
Person editing is improved with better facial identity
preservation across different portrait styles and pose transformations.
Product editing retains identity when used for product posters. Text
editing goes beyond content changes by supporting fonts colors and
materials. The model also support the ControlNet is available with depth maps
edge maps keypoint maps and more. The model is registered under Apache
2.0 which allows open and flexible usage.
Qwen-Image-Edit-2509
addresses the earlier problems with targeted improvements. Multi-image
editing becomes more natural as the model treats inputs as part of a
unified canvas. Single-image consistency is strengthened through better
identity preservation for people and products as well as reliable
formatting for text. ControlNet compatibility provides additional
precision for editing tasks without needing multiple separate tools.
Installation
1. Update ComfyUI from Manager by selecting Update All if you are familiar to it. New user have to install ComfyUI.
2. Download Qwen Image Edit 2509 FP8/BF16 (qwen_image_edit_2509_bf16.safetensors or qwen_image_edit_2509_fp8_e4m3fn.safetensors ) and save it into ComfyUI/models/diffusion_models directory.
BF 16 requires at least 40 GB VRAM and for FP8 its 16 GB.
For low VRAM users, download Qwen Image Edit 2509 gguf and save it into your ComfyUI/models/unet folder. Make sure you have already installed ComfyUI-GGUF custom node by Author-city96.
If not done, install it from the Manager by selecting Custom Nodes Manager option. If you do not know what's GGUF models, you can get detailed overview from our model quantization tutorial.
All the GGUF model details with VRAM usage given below:
2-bit Q2_K 7.06 GB
3-bit Q3_K_S 8.95 GB, Q3_K_M 9.68 GB
4-bit Q4_K_S 12.1 GB, Q4_0 11.9 GB, Q4_1 12.8 GB, Q4_K_M 13.1 GB
5-bit Q5_K_S 14.1 GB, Q5_0 14.4 GB, Q5_1 15.4 GB, Q5_K_M 14.9 GB
6-bit Q6_K 16.8 GB
8-bit Q8_0 21.8 GB
3. All the other models (text encoders, VAE, lora) will be same as used for basic Qwen Image Edit model. You do not need to download again. But if you want then download them :
(a) Download Text encoder (qwen_2.5_vl_7b_fp8_scaled.safetensors) and save it into your ComfyUI/models/text_encoders directory.
(b) Download Vae (qwen_image_vae.safetensors) and save it into ComfyUI/models/vae folder.
(c) Download Lora model (Qwen-Image-Lightning-4steps-V1.0.safetensors) if you want then save it into your ComfyUI/models/loras directory.
4. That's it, just restart and refresh ComfyUI to take effect.
Workflow
1. Download the workflow (Qwen_Image_Edit_2509_Multi_Editing.json) from our Hugging face repository.
 |
| Qwen Image Edit 2509 Workflow |
2. Drag and drop into ComfyUI.
(a) Load multiple images into load Image node. You can use 1-3 images to merge and get the single image. You can enable/ disable the node by right clicking on it and select Bypass.
(b) Load Qwen image edit 2509 model into Unet loader/Load diffusion model loader node.
(c) Load the Vae and text encoders.
(d) Set the KSampler Settings for different model variants:
BF16 variant-
CFG-4
Steps- 50
FP8 variant-
CFG-2.5
Steps- 20
FP8 + lighting lora variant-
CFG-1
Steps- 4
(e) Put the prompt into prompt box.
For Ex- We want to change the clothing of model. We inputted two images- model (image 1) and denim short pants (image2). Here, we used two input images but you can use max three as input.
So, the prompt will be something like-
Replace the skirt in image 1 with denim hot pants in image 2
You need to be expressive so that the qwen image edit 2509 model can understand what type of editing you actually want from it.
(f) Hit Run to execute the workflow.
Test 1 (Clothing Style replacement)
Test 2 (Style/Object transfer)
Here, we want to change the top from image 1 to image 2.
Prompt used: Change and transfer the girl's top in image 2 from image 1.
Test 3 (Pose transfer)
Here, we want to transfer the human pose from image 2 into image 1.
Prompt used: Let the girl in image 1 replicate the pose from image 2.
Qwen-Image-Edit-2509 feels like a strong step forward rather than a minor update. The balance between multi-image capability and single-image consistency gives creators a reliable tool for both professional and personal projects.
The Apache 2.0 license adds confidence for adoption in
open-source and commercial settings. The model shows some limitations
with performance beyond three images but the direction is promising.
