Qwen-Image-Edit 2511 in ComfyUI: Normal and Lightning 4-Step Setup (2026)

Qwen-Image-Edit 2511 is still the edit model many ComfyUI users rely on. Most edit LoRAs on Civitai were trained for it, it keeps faces and products steady across edits, and with the Lightning LoRA it edits in just 4 steps.

This guide sets up both modes in ComfyUI: the normal 40-step model for the best quality and the Lightning 4-step mode for speed. You get every file, folder, node and setting, plus fixes for the red tint and blurry results people hit most.

Why Use Edit 2511 in Late 2026?

Qwen-Image-Edit 2511 came out on 23 December 2025 as an upgrade to 2509. It is a 20B edit model that uses the same text encoder and VAE as the original Qwen-Image.

The newer Qwen-Image 2.1 is smaller and handles up to 10 references, so why bother with 2511? A few good reasons:

• LoRA library: camera-angle, relight, inflation, object-removal and style LoRAs were mostly trained for 2509 and 2511. Edit LoRAs rarely work across versions.
• Consistency: 2511 reduced image drift and keeps people and products steady, which is why it is still popular for product shots and group photos.
• Built-in community LoRAs: Qwen merged some popular community edit effects into the base model.
• Official 4-step Lightning: the LightX2V Lightning LoRA is mature and built into the template.

Still not sure which Qwen edit model to pick? My Qwen Image Edit 2509 vs 2511 vs 2512 vs 2.1 comparison breaks it down.

Normal vs Lightning: What Changes

Lightning is not a separate model. It is a LoRA from LightX2V (ModelTC) that teaches 2511 to finish an edit in 4 steps instead of 40.

Note: at CFG 4 every step runs the model twice (positive and negative). At CFG 1 it runs once. That is why Lightning feels even faster than “40 to 4 steps” suggests.

What You Need: VRAM, RAM and ComfyUI

2511 is a big model, so system RAM matters as much as VRAM. ComfyUI streams parts of the model between RAM and the GPU when the card is too small.

These rows are my estimates from the file sizes and ComfyUI’s offloading, not official figures. If you hit the “Reconnecting” crash, it is usually system RAM running out; see how much RAM you need for local AI.

Any ComfyUI release from the last few months runs 2511 (the template needs v0.4.0 or newer, the INT8 template v0.23.0 or newer). Update anyway, because the template uses new switch nodes. Need a fresh install? See how to install ComfyUI.

Every File and ComfyUI Folder

The diffusion models are in Comfy-Org/Qwen-Image-Edit_ComfyUI, the VAE in Comfy-Org/Qwen-Image_ComfyUI, the text encoder in Comfy-Org/HunyuanVideo_1.5_repackaged and the Lightning files in lightx2v/Qwen-Image-Edit-2511-Lightning.

Warning: the Lightning repo also has a fused FP8 file without “_comfyui” in its name. Users reported it does not load in ComfyUI; the maintainers pointed to the ..._lightning_comfyui.safetensors version instead.

Already have Qwen-Image 2512 or Edit 2509 set up? Then you already have the text encoder and VAE. My Qwen-Image model files guide lists which files each version shares.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── qwen_image_edit_2511_fp8mixed.safetensors
    ├── text_encoders/
    │   └── qwen_2.5_vl_7b_fp8_scaled.safetensors
    ├── vae/
    │   └── qwen_image_vae.safetensors
    └── loras/
        └── Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors

Set Up the Normal Model, Step by Step

1. Update ComfyUI. Portable: update\update_comfyui.bat. Git: git pull and pip install -r requirements.txt. Desktop updates itself.
2. Download the four files: the fp8mixed (or int8) model, the text encoder, the VAE and the Lightning LoRA. Put each in the folder from the table.
3. Open the template. Workflow > Browse Templates > Image > Qwen Image Edit 2511 – Material Replacement (two inputs) or Qwen Image Edit 2511 Int8: Image Edit (one input).
4. Load your picture in the first Load Image node. This is the image you are editing, called “image 1” in prompts.
5. Check the switch. Inside the “Image Edit (Qwen-Image 2511)” subgraph, Enable 4steps LoRA? is off by default, so you get 40 steps at CFG 4.
6. Write the instruction in the positive prompt node and press Run.

A 40-step edit with CFG 4 is the slowest way to run 2511, but it gives the cleanest result. Comfy’s own note says 20 steps at CFG 4 is a fair middle ground.

Turn On Lightning 4-Step Mode

The official template already wires in Lightning. You only flip one switch.

1. Open the subgraph and find the Enable 4steps LoRA? boolean node.
2. Set it to true. Three switch nodes now route the model through the LoRA loader, change steps from 40 to 4 and change CFG from 4 to 1.
3. Check the LoRA name in LoraLoaderModelOnly: Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors, strength 1.0.
4. Run. Compare against a normal-mode result with the same seed.

Lightning in your own workflow

Building from scratch or adding Lightning to an older 2511 graph? Add these yourself:

• LoraLoaderModelOnly straight after Load Diffusion Model, strength 1.0.
• KSampler: steps 4, CFG 1, euler, simple, denoise 1.
• Keep ModelSamplingAuraFlow (shift 3.1) and CFGNorm (strength 1) in the model chain.

Tip: if 4 steps looks soft or plasticky, run the same 4-step LoRA at 8 steps. The LightX2V team confirmed both step counts work, and many users prefer 8 for faces.

The fused Lightning model

Prefer one file? Load qwen_image_edit_2511_fp8_e4m3fn_scaled_lightning_comfyui.safetensors in Load Diffusion Model, bypass the LoRA loader, and use 4 steps at CFG 1. You lose the option to switch back to 40-step mode without loading a second model.

How the Workflow Works (Node by Node)

Knowing the chain helps when you add LoRAs or fix errors. The official subgraph runs in this order:

1. Load Diffusion Model (UNETLoader): the 2511 file, weight dtype default.
2. LoraLoaderModelOnly: the Lightning LoRA, only used when the switch is on. Add style or angle LoRAs after it.
3. ModelSamplingAuraFlow: shift 3.1.
4. CFGNorm: strength 1. It keeps colours from burning at CFG 4.
5. Load CLIP: qwen_2.5_vl_7b_fp8_scaled, type qwen_image.
6. FluxKontextImageScale: resizes your input to a size the model likes, so you do not set width and height yourself.
7. TextEncodeQwenImageEditPlus: one for your instruction, one empty for the negative. Each takes up to three images.
8. FluxKontextMultiReferenceLatentMethod: set to index_timestep_zero on both conditionings.
9. VAEEncode of the scaled image, then KSampler, VAEDecode and Save Image.

The template note says the FluxKontextMultiReferenceLatentMethod nodes are not needed with Comfy’s own files, but may be needed with repackaged models from other uploaders. Keep them; they do no harm.

Editing With 2 or 3 Reference Images

TextEncodeQwenImageEditPlus has three image inputs: image1, image2 and image3. The first is the picture you edit; the others are references you name in the prompt.

• Material swap (the template example): “Change the leather of the sofa in image 1 to the fur material in image 2.”
• Product placement: “Place the bottle from image 2 on the table in image 1, matching the lighting.”
• Outfit swap: “Dress the person in image 1 in the jacket from image 2. Keep the face, pose and background.”
• Group shot: “Put the person from image 2 next to the person in image 1, both smiling at the camera.”

Crop reference images tightly around the subject. Big, busy references slow things down and confuse the model about what to copy. Need more than three references? That is where Qwen-Image 2.1 Edit (up to 10) wins.

GGUF Setup for 12 to 16 GB GPUs

Unsloth’s unsloth/Qwen-Image-Edit-2511-GGUF repo has quants from 7.74 GB to 40.9 GB, with important layers kept at higher precision. Unlike Qwen-Image 2.1, the 2511 architecture (qwen_image) is supported by the original city96 ComfyUI-GGUF node.

1. Install ComfyUI-GGUF:

   cd ComfyUI/custom_nodes
   git clone https://github.com/city96/ComfyUI-GGUF
   pip install --upgrade gguf

   If you already installed the leejet fork for Qwen-Image 2.1, keep that instead; it also loads 2511.

2. Put the .gguf file in ComfyUI/models/unet/. New to GGUF? My ComfyUI-GGUF guide covers the loader nodes in detail.
3. Replace Load Diffusion Model with Unet Loader (GGUF) and connect its MODEL output to the LoRA loader.
4. Keep the normal text encoder, VAE and Lightning LoRA. The LoRA works on GGUF models too.

Running two GPUs? You can put the text encoder on the second card; see my ComfyUI multi-GPU guide.

Best Settings Cheat Sheet

Stacking a style or camera-angle LoRA with Lightning? Start the extra LoRA at 0.7 to 1.0 and keep Lightning at 1.0. If results go strange, try 8 steps before lowering strengths.

Edit Prompts That Work

2511 follows short, direct instructions best. Say what to change, then what to keep.

• Change one thing: “Change the car colour to matte black. Keep everything else the same.”
• Change the scene: “Turn this daytime street into a rainy night with neon reflections on the road.”
• Edit text: “Replace the text on the shop sign with “FRESH BREAD”, same font style.”
• Style: “Convert this image to a pop art poster style.” (the Int8 template’s example)
• Remove: “Remove the person in the background on the left and fill in the wall naturally.”

Write prompts in English or Chinese. If a result changes too much, add “keep the face, pose and background unchanged” and use the normal 40-step mode, which follows “keep” instructions more strictly than Lightning.

Troubleshooting

Red or orange tint on Lightning results

This usually means you are running a 2509 workflow with 2511. Use the 2511 template, or add FluxKontextMultiReferenceLatentMethod (index_timestep_zero) on both positive and negative conditioning.

Lightning output is blurry or burnt

CFG is not 1, or you loaded a Lightning LoRA for another model. Use the 2511 Lightning LoRA (not the 2509 or Qwen-Image Lightning V2.0 files) and set CFG to 1.

The fused FP8 Lightning file will not load

Download the file ending in _lightning_comfyui.safetensors. The other fused file is for Diffusers and LightX2V.

“Reconnecting” or ComfyUI crashes mid-run

System RAM ran out while offloading the 20 GB model. Close other apps, raise the Windows page file, or switch to GGUF Q4_K_M.

Out of memory on the GPU

Use a smaller GGUF, keep input images near 1 megapixel and remove the third reference image. More tips in my CUDA out of memory fix.

Mat1 and mat2 shapes cannot be multiplied

The text encoder is wrong. 2511 needs qwen_2.5_vl_7b_fp8_scaled with type qwen_image, not the Qwen3-VL encoder from Qwen-Image 2.1.

The face changes after the edit

Turn Lightning off for that edit, add “keep the face identical” to the prompt and crop references tightly. Identity is where 40 steps beats 4 steps most clearly.

My Take

If you are starting fresh in November 2026, Qwen-Image 2.1 Edit is lighter and newer. But if you use edit LoRAs, or you need the steady product and character consistency 2511 is known for, keep 2511 installed.

Set the template up once with the Lightning LoRA in place. Flip the switch on for quick tests and batches, then flip it off for the final 40-step edit. On a 16 GB card with plenty of system RAM, that combination is what makes a 20B editor practical for everyday use.

FAQ

What is Qwen-Image-Edit 2511 Lightning?

It is a LoRA from LightX2V that lets Qwen-Image-Edit 2511 finish an edit in 4 steps at CFG 1 instead of 40 steps at CFG 4.

Where do I put the Lightning LoRA?

In ComfyUI/models/loras/. Load it with LoraLoaderModelOnly at strength 1.0, or flip the “Enable 4steps LoRA?” switch in the official template.

Can I use the 2509 Lightning LoRA with 2511?

Use the 2511 version. Edit LoRAs are version-specific, and mixing them causes colour shifts and blur.

Does the 4-step LoRA work at 8 steps?

Yes. The LightX2V team said both 4 and 8 steps work, and 8 steps often gives cleaner faces.

Can Qwen-Image-Edit 2511 run on 16 GB VRAM?

Yes, with the fp8mixed or int8 file and ComfyUI offloading (64 GB RAM recommended), or with the GGUF Q4_K_M or Q5_K_M file.

Do 2511, 2509 and Qwen-Image 2512 share files?

Yes. They all use qwen_2.5_vl_7b_fp8_scaled and qwen_image_vae. Only the diffusion model and LoRAs change.

How many images can 2511 combine?

Up to three in one edit: the image you edit plus two references.

Is Qwen-Image-Edit 2511 better than Qwen-Image 2.1 Edit?

Not overall. 2.1 is newer, lighter and takes more references, but 2511 has far more edit LoRAs and very steady identity.

Sources

• ComfyUI docs: Qwen-Image-Edit-2511 tutorial
• Comfy-Org workflow templates (Qwen Image Edit 2511 and 2511 Int8 JSON and notes)
• lightx2v/Qwen-Image-Edit-2511-Lightning and its 8-step and ComfyUI file discussion
• Comfy-Org/Qwen-Image-Edit_ComfyUI
• unsloth/Qwen-Image-Edit-2511-GGUF
• city96/ComfyUI-GGUF