Qwen Image Edit 2509 in ComfyUI: 3-Image Compose (Q5 GGUF)

I am using Qwen Image Edit 2509 inside ComfyUI. In my tests it beats Nano Banana for real world edits. The blend looks natural and identity stays tight. You can run it on a small GPU with Q5 GGUF and still get clean results. Sources for the model and the ComfyUI build are linked below.

Files you need

Put each file in the right ComfyUI folder. If your install uses custom paths, point the nodes there.

Base model

• qwen_image_edit_2509_fp8_e4m3fn.safetensors or qwen_image_edit_2509_bf16.safetensors
  Folder: ComfyUI/models/diffusion_models Hugging Face+1

GGUF for low VRAM

• Use a Q5 build. Example Qwen-Image-Edit-2509-Q5_K_M.gguf
  Folder: wherever your GGUF loader expects it. Common paths are ComfyUI/models/diffusion_models or ComfyUI/models/diffusion_models Hugging Face

LoRA for speed

• Lightning 4 step and 8 step for Qwen Image
  Folder: ComfyUI/models/loras Hugging Face+1

Text encoders

• CLIP and T5 that ship with the Comfy-Org split files
  Folder: ComfyUI/models/clip Hugging Face+1

VAE

• qwen_image_vae.safetensors
  Folder: ComfyUI/models/vae Hugging Face+1

What this workflow does

You can edit with up to three images

• Image 1 is the person
• Image 2 is the background or outfit reference
• Image 3 is a pose reference if you need it

I added a Resolution Master and fed all three images into it. Same size in equals cleaner output. I also drive three Resize nodes from the same master so every branch matches. Multi image editing is a main feature in the 2509 release.

I made a Select Group with three switches

• Generate with one image
• Generate with two images
• Generate with three images
  Turn on only the one you need.

I also dropped an Upscale group. Pick Yes or No. If you use one image, route Result Image 1. For two images, route Result Image 2. For three images, route Result Image 3.

Model setup that works on a small GPU

If you have Good VRAM, use FP8 or BF16 from Comfy-Org. Hugging Face

If you are on low VRAM, use Q5 GGUF. From my tests, skip Q2 to Q4. Start at Q5. Move to Q6 or Q8 if you have headroom. Quantized 2509 builds are here. Hugging Face

Lightning loader

• 4 step for low RAM
• CFG is 1

Part 1. One image tests on Q5

Turn on the one image group. Upload one photo of a woman.

Turn the face
Prompt: turn the face toward the background
Steps 4
CFG 1

Result comes in a few seconds on Q5. The face turns and the identity holds. Community tests agree that the 2509 edit build improved single image consistency.

Victory sign
Prompt: victory sign

Older edit models often fudge fingers. Here the hand renders with the right count. Recent posts show better hand accuracy on 2509.

Blue hair
Prompt: make the hair color blue

Edges stay clean. Blend looks natural. Qwen’s edit branch is built to preserve details while changing appearance.

Clothes and background
Prompt: yellow t shirt and jeans
Prompt: make the background a park with grass

Outfit swaps and background swaps are stable in 2509. If background leaks, raise Steps by one or two and try again.

Glasses and expression
Prompt: wear glasses
Prompt: make an angry expression

Small edits like these are straightforward on 2509 and show up in public demos and spaces.

Two image compose

Turn on the two image group and bypass the one image group

Image 1 is the person.
Image 2 is the outfit reference.

Example outfit in Image 2
Brown overcoat. Plain scarf. Black turtleneck. Tailored black trousers. Knee high black boots. Brown purse.

Prompt style I like
Role A is the person. Role B is the outfit reference. Keep from A the face, hair, body proportion, pose baseline, identity, stance, background. Dress her in B.

Run Q5 with 4 steps. Result matches the coat and scarf. If pose shifts, add same pose to the prompt and run again. Multi image compose is a headline upgrade in 2509.

Second subject test

I switch Image 1 to a woman in a white dress. I keep the same Image 2. Clothing applies. One miss. The purse is not brown. I add brown handbag to the prompt.

Next run, the purse is brown and matches the reference texture.

Put the woman from Image 1 on the blue chair in Image 2. Keep her clothes and face same as Image 1. Keep the beach background. Works as expected. The wrist band and small details survive.

Three image control

Goal
Use Image 1 as the person, Image 2 as the background, Image 3 as the pose reference. Keep identity from Image 1. Keep the background from Image 2. Copy the pose from Image 3.

Example
Replace an anime girl in Image 2 but keep the same pose. First run misses the posture. I enable three images and load the same anime frame as Image 3 to copy pose. Prompt line get the same pose from image 3. Now the posture matches and the background stays the same.

Add one more line to keep extra identity
Keep the same hair pinky tattoo body proportion and black ruffled dress from image 1. Keep the same pose from image 3.

The final blend looks clean.

One click upscale

Turn on the Upscale group at the end. Pick your preferred upscaler. Many community spaces ship ready examples for 2509. Upscaling helps when you use small safe sizes on a low VRAM GPU.