Skip to content
,

How to Run Qwen-Image Locally (2026): Create and Edit Images in ComfyUI, Every Setup Explained

7–11 minutes
Excited silver-haired woman in a green bomber jacket on a neon-lit city street holding a QWEN card, with the hook text Free AI Images on Your PC

Qwen-Image is one of the best open image model families you can run at home. It is excellent at readable text inside images, follows long prompts well, and its edit models can change clothes, swap backgrounds, remove objects and combine several reference pictures.

The catch is that there are now several versions, two kinds of models (create and edit), normal and turbo modes, and different files for each.

This guide brings everything together in one place.

You will learn which Qwen-Image version to use, how to set up all four local setups in ComfyUI (create normal, create turbo, edit normal, edit turbo), exactly which files go in which folder, how much VRAM each needs, and how to fix the common errors.

Each section links to our detailed step-by-step posts. Facts are checked against the Comfy-Org repos, the official ComfyUI templates and the ComfyUI-GGUF code in November 2026 (ComfyUI is at v0.46.0).

TL;DR

  • Latest version: Qwen-Image 2.1 does both creating and editing. Older Qwen-Image-Edit 2511 is still popular for edits with its 4-step Lightning LoRA.
  • Templates: ComfyUI > Workflow > Browse Templates > Image: “Qwen Image 2.1: Text to Image” (25 steps), “Qwen Image 2.1 Turbo: Text to Image” (8 steps), “Qwen Image 2.1: Image Edit” and “Qwen Image 2.1 Turbo: Image Edit”.
  • Update first: ComfyUI 0.37.2+ for the normal 2.1 template, 0.39.0+ for Turbo.
  • 16 GB card (2.1): int8 model (7.26 GB) in diffusion_models, qwen3vl_8b_int8_convrot (9.35 GB) in text_encoders, qwen_image_2.1_vae_bf16 in vae.
  • 8-12 GB: GGUF Q4 to Q8 in models/unet with the leejet/ComfyUI-GGUF fork (city96’s node still lacks the 2.1 architecture).
  • Do not mix files: 2.1 uses a Qwen3-VL encoder and its own VAE; older Qwen-Image uses qwen_2.5_vl_7b and qwen_image_vae.
  • Licence: Qwen-Image 2.1 uses the Qwen Research License (non-commercial terms); check before commercial use.

Which Qwen-Image Version Should You Use?

Version Does Text encoder / VAE Best for
Qwen-Image 2.1 Create and edit (up to 10 references) Qwen3-VL 8B / qwen_image_2.1_vae Most people in 2026: best quality, int8 and GGUF files fit 8-16 GB
Qwen-Image-Edit 2511 Edit only qwen_2.5_vl_7b / qwen_image_vae Material and outfit swaps with the 4-step Lightning LoRA; 24 GB cards or GGUF
Qwen-Image-2512 Create only qwen_2.5_vl_7b / qwen_image_vae Existing workflows and LoRAs built for the older base
Qwen-Image-Edit 2509 Edit only qwen_2.5_vl_7b / qwen_image_vae Older LoRAs only

If you are starting fresh, use Qwen-Image 2.1 for everything. For a detailed side-by-side, see our Qwen-Image Edit 2509 vs 2511 vs 2512 vs 2.1 comparison.

VRAM Guide: Pick Your Setup

Your GPU Create Edit
8 GB 2.1 GGUF Q4 (4.2 GB) + w4a8 encoder (6.31 GB) 2.1 GGUF Q4_K_M + w4a8 encoder; keep 1024 px
12 GB 2.1 GGUF Q6_K or Q8_0 (7.6 GB) 2.1 GGUF Q6_K/Q8_0; or 2511 GGUF Q4_K_M (12.8 GB) with offloading
16 GB 2.1 int8 (7.26 GB) + int8 encoder (9.35 GB) 2.1 Turbo int8 + int8 encoder
24 GB+ 2.1 int8 or bf16 2.1, or Edit 2511 fp8mixed/int8 (20.5 GB) + Lightning LoRA

ComfyUI offloads to system RAM when VRAM runs out, so 32 GB of RAM makes all of these smoother. Not sure what a quant level means? Our GGUF quantization guide explains Q4 vs Q6 vs Q8.

Before You Start

  1. Update ComfyUI. The normal 2.1 template needs v0.37.2 or newer and the Turbo template needs v0.39.0 or newer. The current release is v0.46.0. New to ComfyUI? Start with our ComfyUI beginner guide.
  2. Update your GPU driver and make sure you have 30-50 GB of free disk space.
  3. Open the template browser: Workflow > Browse Templates > Image, and search “Qwen”.
  4. Let the template tell you the files. When you open a template, ComfyUI lists missing models with download links. Put each file in the folder shown.

Setup 1: Create Images (Normal, 25 Steps)

  1. Open the template “Qwen Image 2.1: Text to Image”.
  2. Download qwen_image_2.1_int8_convrot.safetensors (7.26 GB) into ComfyUI/models/diffusion_models/.
  3. Download qwen3vl_8b_int8_convrot.safetensors (9.35 GB) into ComfyUI/models/text_encoders/.
  4. Download qwen_image_2.1_vae_bf16.safetensors (0.68 GB) into ComfyUI/models/vae/.
  5. Restart ComfyUI or press R to refresh, then select the files in the loader nodes.
  6. Use euler, simple, CFG 1, 25 steps at 1024 x 1024. (The official Diffusers pipeline uses 40-50 steps; more steps add a little detail at the cost of time.)
  7. Write your prompt and click Run.

Qwen-Image shines with text: put the exact words in quotes, for example a bakery shop window with a sign that says “Fresh Bread Daily”.

Setup 2: Create Images (Turbo, 8 Steps)

  1. Open “Qwen Image 2.1 Turbo: Text to Image” (needs ComfyUI 0.39.0+).
  2. Download qwen_image_2.1_turbo_int8_convrot.safetensors (7.26 GB) into models/diffusion_models/. The text encoder and VAE are the same as Setup 1.
  3. Keep the template’s settings: euler, CFG 1, 8 steps with its fixed sigma list (ManualSigmas node). Do not just change the step count on the normal workflow; the sigma list matters.
  4. Run. Turbo is roughly three times faster than 25 steps with similar quality for most prompts.
Other fast options Steps Notes
Official Turbo LoRA on the base model 8 Handy with GGUF base files
Pruna P-Q-Image-2.1 LoRA 6 Community distill LoRA
Lightning v2 RL 4 Diffusers checkpoint with a custom sigma list; v1 Lightning LoRA (67 MB) works in ComfyUI

Setup 3: Edit Images (Normal)

  1. Open “Qwen Image 2.1: Image Edit” (25 steps).
  2. Use the same 2.1 model, encoder and VAE as Setup 1. The encoder’s vision part is what lets the model “see” your picture.
  3. Load your photo into the first image input (image_1). This is the picture that gets edited.
  4. Add extra references if needed (up to 10). In the prompt, refer to them as <image2>, <image3> and so on, for example put the jacket from <image2> on the woman in the photo.
  5. Settings: 25 steps, CFG 1, euler/simple. Keep 1024 px while testing.
  6. Run, then refine the prompt with one change at a time.

There is also a ready-made “Remove Background: Qwen Image 2.1” template. Full walkthrough with examples: Qwen Image 2.1 Edit in ComfyUI. To erase objects, see our Qwen-Image remove object LoRA guide.

Only edit photos you own or have permission to use, and never edit real people to mislead or embarrass them.

Setup 4: Edit Images (Turbo / Lightning)

Option A: Qwen-Image 2.1 Turbo Edit (16 GB)

  1. Open “Qwen Image 2.1 Turbo: Image Edit”.
  2. Load qwen_image_2.1_turbo_int8_convrot.safetensors with the int8 encoder and 2.1 VAE.
  3. Use 8 steps, CFG 1, euler.

Option B: Qwen-Image-Edit 2511 + Lightning (24 GB, or GGUF)

  1. Open “Qwen Image Edit 2511 Int8: Image Edit” or “Qwen Image Edit 2511 – Material Replacement”.
  2. Model: qwen_image_edit_2511_int8_convrot.safetensors or qwen_image_edit_2511_fp8mixed.safetensors (20.5 GB) in diffusion_models.
  3. Encoder: qwen_2.5_vl_7b_fp8_scaled.safetensors (9.38 GB) in text_encoders. VAE: qwen_image_vae.safetensors (254 MB) in vae.
  4. LoRA: Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors (850 MB) in models/loras, strength 1.0.
  5. Flip the template’s “Enable 4steps LoRA?” switch on, then use 4 steps (or 8) at CFG 1. Without the LoRA, use 40 steps, CFG 4, euler, simple, shift 3.1.

Step by step with screenshots: Qwen-Image-Edit 2511 Lightning in ComfyUI.

Low VRAM: GGUF on 8-12 GB Cards

  1. Install the leejet/ComfyUI-GGUF fork through ComfyUI Manager or git clone into custom_nodes. Its loader includes the qwen_image21 architecture. The original city96/ComfyUI-GGUF still lists only qwen_image and fails on 2.1 files with “Unknown model architecture!”.
  2. Download a Qwen-Image 2.1 GGUF: Q4 (about 4.2 GB) for 8 GB, Q6_K (about 6-6.3 GB) or Q8_0 (about 7.6 GB) for 12 GB. Put it in ComfyUI/models/unet/.
  3. Use a smaller encoder: qwen3vl_8b_w4a8.safetensors (6.31 GB) in text_encoders.
  4. Replace the template’s model loader with Unet Loader (GGUF) and keep everything else.
  5. For turbo speed with GGUF, add the official Turbo LoRA and use 8 steps with the template’s ManualSigmas.

For older Qwen-Image and Edit 2511 GGUF files, city96’s node works fine. Full details: Qwen-Image 2.1 GGUF on 8 GB and 12 GB.

File and Folder Cheat Sheet

File Size Folder Used by
qwen_image_2.1_int8_convrot.safetensors 7.26 GB models/diffusion_models 2.1 create/edit
qwen_image_2.1_turbo_int8_convrot.safetensors 7.26 GB models/diffusion_models 2.1 Turbo
qwen3vl_8b_int8_convrot.safetensors 9.35 GB models/text_encoders 2.1
qwen3vl_8b_w4a8.safetensors 6.31 GB models/text_encoders 2.1 low VRAM
qwen_image_2.1_vae_bf16.safetensors 0.68 GB models/vae 2.1
Qwen-Image 2.1 GGUF (Q4-Q8) 4.2-7.6 GB models/unet 2.1 with leejet node
qwen_image_edit_2511_int8_convrot / fp8mixed 20.5 GB models/diffusion_models Edit 2511
qwen_2.5_vl_7b_fp8_scaled.safetensors 9.38 GB models/text_encoders All pre-2.1 versions
qwen_image_vae.safetensors 254 MB models/vae All pre-2.1 versions
Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16 850 MB models/loras Edit 2511 turbo

More detail on every file: Qwen-Image model files explained. Official downloads live in the Comfy-Org repos on Hugging Face (Qwen-Image-2.1, Qwen-Image_ComfyUI and Qwen-Image-Edit_ComfyUI).

Prompt Tips for Qwen-Image

  • Put text in quotes and say where it goes: a poster with the title “SUMMER SALE” in bold white letters at the top.
  • Describe layout in order: subject, setting, lighting, camera, style.
  • For edits, say what stays the same: change the sofa to green velvet, keep everything else identical.
  • One change per edit pass gives cleaner results than five changes at once.
  • CFG stays at 1 for 2.1 and turbo setups; raising it usually burns the image.

Troubleshooting

Problem Likely cause Fix
“Unknown model architecture!” with a 2.1 GGUF Using city96’s ComfyUI-GGUF Install the leejet/ComfyUI-GGUF fork
Template missing from the browser Old ComfyUI Update to 0.37.2+ (normal) or 0.39.0+ (Turbo)
Garbled, noisy or black images Mixed 2.1 and older encoder/VAE files Use the Qwen3-VL encoder and 2.1 VAE with 2.1 models only
Edit ignores the input photo Text encoder without vision part, or GGUF encoder missing its mmproj Use the full encoder; keep the matching mmproj file next to a GGUF encoder
Turbo results look washed out Wrong sigmas or CFG Use the Turbo template’s ManualSigmas and CFG 1
Out of memory Model plus encoder too big Switch to GGUF and the w4a8 encoder, lower resolution
2511 Lightning images look overcooked LoRA plus high CFG CFG 1 and 4-8 steps with the LoRA

Still stuck? Our ComfyUI black image fix covers VAE and precision problems in depth.

All Our Qwen-Image Guides

FAQ

What is the latest Qwen-Image version?

Qwen-Image 2.1, which handles both text-to-image and editing with up to 10 reference images.

Can I run Qwen-Image on 8 GB VRAM?

Yes. Use a Qwen-Image 2.1 GGUF Q4 with the w4a8 text encoder and the leejet ComfyUI-GGUF node, at 1024 x 1024.

What is the difference between normal and Turbo?

Normal uses 25 steps; Turbo uses a distilled model with 8 steps and a fixed sigma list, so it is about three times faster with similar quality.

Which text encoder does Qwen-Image 2.1 use?

A Qwen3-VL 8B encoder such as qwen3vl_8b_int8_convrot.safetensors. Older versions use qwen_2.5_vl_7b, and the two are not interchangeable.

Why does city96 ComfyUI-GGUF fail with Qwen-Image 2.1?

Its loader does not include the qwen_image21 architecture. The leejet fork does.

Should I use Qwen-Image 2.1 Edit or Edit 2511?

2.1 for most edits and lower VRAM. 2511 with its Lightning LoRA is still great for material and outfit swaps if you have 24 GB or use GGUF.

Can I use Qwen-Image commercially?

Qwen-Image 2.1 is released under the Qwen Research License with non-commercial terms. Read the licence on the model card before any commercial use.

Sources

Free weekly email

Get new AI guides every week

One short weekly email with our newest local AI, ComfyUI and AI tool guides. No spam, no pop-ups, just the good stuff.

Free. Unsubscribe anytime in one click. We only use your email to send this newsletter. See our privacy policy.

Excited silver-haired woman in a green bomber jacket on a neon-lit city street holding a QWEN card, with the hook text Free AI Images on Your PC