Qwen Image Edit 2511 in ComfyUI: a beginner guide

Qwen-Image-Edit 2511 is an instruction-based image editor. You give it a picture and a sentence, and it changes what you asked for while keeping the rest. For anyone building a consistent AI character, it is the most useful model in ComfyUI right now, because it keeps faces remarkably well.

This guide takes you from a fresh ComfyUI install to your first edit, with the exact settings we use in the AI Character Kit. It assumes you have used ComfyUI a little, but not much more.

What does Qwen Image Edit 2511 actually do?

Think of it as “Photoshop by sentence”. Some things it is good at:

• Putting the same person in a new place, outfit or pose
• Changing the art style (anime, 3D cartoon, oil painting) while keeping the face shape
• Removing or adding objects
• Fixing lighting or time of day

The model is from the Qwen team, with an official page at Qwen/Qwen-Image-Edit-2511. For ComfyUI you want the repackaged files from Comfy-Org, which are ready to drop into the right folders.

What hardware do you need?

The fp8 version of the model runs well on a 24 GB NVIDIA card such as an RTX 3090 or 4090. In our tests, peak VRAM for image work was about 17.4 GB. You also want 32 GB of system RAM. If you don’t have a card like that, you can rent one by the hour; see running ComfyUI on Vast.ai cheaply.

Update ComfyUI first. The portable build has update/update_comfyui.bat. For a git install, run git pull in the ComfyUI folder, then pip install -r requirements.txt. Older versions are missing the Qwen edit text encode node.

Which files do you download, and where do they go?

You need four files. All are free.

File	Folder	Source
qwen_image_edit_2511_fp8mixed.safetensors	models/diffusion_models	Comfy-Org/Qwen-Image-Edit_ComfyUI
qwen_2.5_vl_7b_fp8_scaled.safetensors	models/text_encoders	Comfy-Org/Qwen-Image_ComfyUI
qwen_image_vae.safetensors	models/vae	Comfy-Org Qwen Image pages
Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors	models/loras	lightx2v/Qwen-Image-Edit-2511-Lightning

The Lightning LoRA is optional but worth it. It cuts the steps from 20 or more down to 4, which takes a typical edit from about a minute to about 15 seconds on a 3090 Ti.

After copying the files, restart ComfyUI or press R in the interface to refresh the model lists. The official ComfyUI Qwen Image Edit 2511 tutorial has a template workflow if you want to compare.

How do you wire the workflow?

Here is the node chain, from left to right. If a node name is unfamiliar, double click the canvas and type it.

1. Load Diffusion Model (UNETLoader): pick qwen_image_edit_2511_fp8mixed, weight dtype default.
2. LoraLoaderModelOnly: pick the Lightning 4 step LoRA, strength 1.
3. ModelSamplingAuraFlow: shift 3.1.
4. CFGNorm: strength 1.
5. Load CLIP: pick qwen_2.5_vl_7b_fp8_scaled, type qwen_image.
6. Load VAE: pick qwen_image_vae.
7. Load Image: your character image.
8. ImageScaleToTotalPixels: lanczos, 1 megapixel. This keeps memory and speed predictable whatever size your input is.
9. TextEncodeQwenImageEditPlus for the positive prompt, connected to the CLIP, the VAE and the scaled image.
10. A second TextEncodeQwenImageEditPlus for the negative prompt. Leave its text empty.
11. VAE Encode the scaled image to get the starting latent.
12. KSampler, then VAE Decode and Save Image.

The kit’s workflow also routes both conditionings through a FluxKontextMultiReferenceLatentMethod node set to index_timestep_zero, which helps the model treat the input image as a reference.

What are the right KSampler settings?

Setting	With Lightning LoRA	Without LoRA
Steps	4	20 or more
CFG	1	about 4
Sampler	euler	euler
Scheduler	simple	simple
Denoise	1	1

The most common beginner mistake is leaving CFG at the default of 7 or 8 with the Lightning LoRA. You get burnt, oversaturated images. With the LoRA, CFG must be 1. If you bypass the LoRA for maximum quality, raise steps to 20 or more and CFG to about 4.

How do you write edit prompts that work?

Edit prompts are instructions, not descriptions. Tell the model what to keep and what to change. This structure works reliably for people:

Show the same [woman/man] from the image, same face, same hair and [unique mark],
now [action] [place] [time/light] in [outfit], [camera style], natural skin texture

A few tested examples, all starting with “Show the same woman from the image, same face, same hair and same nose stud, now…”:

...standing on a rainy neon street at night in a black leather jacket, holding a
clear umbrella, reflections on the wet road, candid photo

...taking a mirror selfie in a modern gym in a sage green sports bra and matching
leggings, phone in hand, bright overhead light

...sitting in a cozy coffee shop holding an oat latte, wearing a camel
turtleneck, warm afternoon light, candid photo

For a style change, keep the identity words and describe the look:

Turn the woman into a 3D animated movie character, same face, same hair, same
outfit, soft studio lighting, big expressive eyes

Why does the face still drift sometimes?

Usually one of these:

• You edited an edited image. Always go back to the original base image. Every generation adds a little drift.
• You changed too much at once. New place plus new outfit is fine. Add a new haircut and a new art style and identity starts to slip.
• You didn’t name the small details. Jewellery, tattoos and ribbons disappear unless you mention them: “keep the small gold nose stud”.
• The face in the input is small. Waist up works far better than full body.
• Bad luck. Reroll the seed before you rewrite the prompt.

If skin looks plastic, remove words like “perfect”, “flawless” or “8k” and add “natural skin texture with visible pores, slightly grainy”.

Where does your input image come from?

Any clear photo of a fictional character works. In the kit, the base image comes from Krea 2 Turbo (8 steps, CFG 1, euler / simple, 896 x 1344), which gives realistic skin. The full method is in how to keep the same face without LoRA training.

Please only edit images of fictional characters or of people who gave you clear permission. Editing real people into scenes they were never in is how deepfakes start.

What can you do with the edited scenes?

Each scene can be animated. Feed it into Wan 2.2 image to video for a 5 second clip; the Wan 2.2 settings guide covers that. Or use it as the reference for a character swap with Wan Animate 2.

If you would like this exact workflow ready to load, with the settings above, a sample character and 64 tested prompts, it is workflow 02 in the AI Character Kit.