Qwen Image Edit is one of the most capable open-weights image models for instruction-based editing. It combines 20 billion parameters with native ControlNet support, multi-image editing, and excellent text rendering, capabilities previously restricted to closed-source APIs.
Qwen AI was developed by Alibaba Cloud, the group behind the broader Qwen family of language and vision models. This guide covers everything from understanding the model to downloading files, installing in ComfyUI, and getting the best results from your edits.
Qwen Image: An Overview
20B Parameters, Text Rendering, and Editing
Qwen-Image is a 20B parameter MMDiT (Multimodal Diffusion Transformer) model released under the Apache 2.0 license. Its architecture combines two large components: a 20.4B DiT Transformer that constructs the image, and an 8.3B Qwen2.5-VL text encoder that processes the prompt.
Multimodal Diffusion Transformer (MMDiT) is a generative neural network architecture that combines a transformer with a diffusion (denoising) process to process and align multiple data in a single framework.
This dual-encoding mechanism enables the editing module to balance semantic consistency and visual fidelity, which is why the model produces instruction-following edits that larger but simpler architectures struggle with.
Qwen-Image's standout strength is complex text rendering. It generates images with multi-line layouts and paragraph-level content in both Chinese and English, a capability most other open-weight models handle poorly.
Qwen Image Edit vs. Flux Kontext: Key Differences
Qwen Image is significantly larger than Flux at 20.4B vs. 12B parameters. It uses a Qwen2.5-VL vision-language model as its text encoder, giving it strong bilingual understanding that Flux models don't match.
| Capability | Qwen Image Edit | Flux Kontext Dev |
|---|---|---|
| Multi-image editing (up to 3 inputs) | Strong | Limited |
| Text rendering in images | Excellent (bilingual) | Good (English) |
| Instruction-based editing precision | Excellent | Good |
| Generation speed (inference steps) | 20–40 steps (FP8) | 28 steps |
| Native ControlNet support | Built-in (depth, edge, keypoint) | Via adapters |
| Min VRAM (quantized) | 8GB (Q2_K GGUF) | 6–8GB (GGUF) |
| License | Apache 2.0 | Non-commercial |
Where Flux Kontext leads on generation speed, Qwen Image Edit is the stronger tool for tasks that require combining multiple reference images, editing text within an image, or applying precise object-level changes that need to be preserved across identity and composition.
What's New in Qwen Image Edit 2509 and 2511?
2509 (September 2025)
Qwen-Image-Edit-2509 introduced multi-image editing support, covering person+person, person+product, and person+scene combinations. It also added native ControlNet support with built-in depth maps, edge maps, and keypoint maps.
Three areas saw major improvements:
- Person editing - better facial identity preservation
- Product editing - identity retention across styles
- Text editing - font type, color, and material changes
2511 (December 2025)
Qwen-Image-Edit-2511 was released on December 26, 2025, building on 2509 with improved multi-image editing consistency, multi-angle view generation from a single reference image (re-shoot from any camera angle), and refined face identity preservation across pose and style transformations. Native LoRA support is built-in, so community LoRAs work without additional calibration nodes.
The ComfyUI workflow structure is largely the same between versions. The primary change is the checkpoint filename. Use qwen_image_edit_2511_fp8_e4m3fn.safetensors instead of the 2509 variant if you want the December 2025 update. The text encoder and VAE files are shared across versions.
Qwen-Image VRAM Requirements: What GPU Do You Need?
VRAM Requirements by Precision
| Format | File Size | Min VRAM | Notes |
|---|---|---|---|
| BF16 | 40.9GB | 48GB+1 | Full precision, best quality |
| FP8 | 20.4GB | 24GB1 | Recommended for 24GB GPUs. Minimal quality loss. |
| Q8_0 GGUF | 21.8GB | 24GB2 | Near-original quality |
| Q4_K_M GGUF | 13.1GB | 14GB2 | Excellent quality, widely recommended |
| Q3_K_M GGUF | 9.8GB | 10GB2 | Good quality for budget GPUs |
| Q2_K GGUF | 7.5GB | 8GB2 | Lower quality, extreme constraints |
| 1 Includes the text encoder and VAE loaded in VRAM (~8GB). 2 Text encoder and VAE offloaded to system RAM. |
|||
Running Qwen Image Edit on 8GB, 16GB, and 24GB GPUs
On 8GB, you are limited to Q2_K or Q3_K_S quantizations with the Lightning LoRA enabled. Results are usable but noticeably softer than higher-precision variants.
At 16GB, FP8 is the sweet spot: near-full quality without GGUF quantization. The unquantized model needs 24GB or more (such as an RTX 4090) along with 64GB of system RAM for optimal performance.
To overcome local hardware limitations, Thunder Compute offers RTX A6000 GPUs with 48GB of VRAM at $0.35/hr.
How to Install Qwen Image Edit in ComfyUI
Update ComfyUI to the latest version before downloading anything. Native support for Qwen Image Edit 2511 was added in December 2025, and any build predating that could show red-node errors on load.
Downloading the Model Files
All files are available for free on Hugging Face. Download each file and place it in the directory shown.
| File | Size | Destination Directory |
|---|---|---|
| qwen_image_edit_2509_fp8_e4m3fn.safetensors | ~21GB | ComfyUI/models/diffusion_models/ |
| qwen_image_edit_2509_bf16.safetensors (optional) | ~41GB | ComfyUI/models/diffusion_models/ |
| qwen_2.5_vl_7b_fp8_scaled.safetensors | ~8GB | ComfyUI/models/text_encoders/ |
| qwen_image_vae.safetensors | ~335MB | ComfyUI/models/vae/ |
| Qwen-Image-Lightning-4steps-V1.0.safetensors (optional) | ~800MB | ComfyUI/models/loras/ |
| For GGUF variants, download from community repositories on Hugging Face and place in ComfyUI/models/unet/. Install the ComfyUI-GGUF custom node by city96 first. | ||
For low-VRAM setups, download a GGUF variant instead and place it in ComfyUI/models/unet/. You will also need the ComfyUI-GGUF custom node by city96, installable through ComfyUI Manager.
Directory Structure
The correct directory layout for model files is as follows:
ComfyUI/
└── models/
├── text_encoders/
│ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
├── diffusion_models/
│ └── qwen_image_edit_2509_fp8_e4m3fn.safetensors
├── unet/
│ └── Qwen-Image-Edit-2509-Q4_K_M.gguf (GGUF path)
├── loras/
│ └── Qwen-Image-Lightning-4steps-V1.0.safetensors (optional)
└── vae/
└── qwen_image_vae.safetensors
Placing files in the wrong directories is the most common setup mistake. After placing all files, do a full ComfyUI restart.
Loading the Official Workflow
The official workflow JSON files are hosted on Comfy-Org GitHub and the ComfyUI examples documentation page. Drag the workflow image or JSON file directly onto the ComfyUI canvas to load it. ComfyUI decodes the embedded metadata and restores the full node graph automatically.
Prefer a simpler interface? Forge Neo is a traditional web UI alternative to ComfyUI that works with Qwen.
How to Use Qwen Image Edit in ComfyUI
Single-Image Editing: Prompts, Steps, and Settings
Load your source image into the Load Image node, write your edit instruction in the positive prompt field, and add a short negative prompt for unwanted elements such as blurriness or watermarks.
Recommended KSampler settings by variant:
- BF16: CFG 4.0, 40 steps
- FP8: CFG 4.0, 20 steps
- FP8 + Lightning LoRA: CFG 1.0, 4 steps
Using the wrong settings for your variant is the most common cause of blocky or poor-quality output. The Lightning LoRA setting (CFG 1.0, 4 steps) is a different inference regime. Mixing it with standard settings degrades quality significantly.
Practical Editing Use Cases with Example Prompts
Qwen Image Edit handles a wide range of tasks that other open-weight models struggle with. Here are the most reliable use cases with effective prompt patterns.
Changing text in an image: "Replace the text on the sign that reads 'SALE' with 'OPEN'. Keep the font style, color, and size consistent with the original."
Swapping outfit or clothing: "Change the jacket from red to navy blue. Keep the person's face, hair, and background unchanged."
Background replacement: "Replace the background with a modern office environment, maintaining the subject in the foreground with consistent lighting."
Style transfer: "Convert this photo to a Studio Ghibli-style illustration, maintaining the subject's features and pose."
Product poster editing: "Place this product in a flat lay composition on a white marble surface with soft natural lighting and a simple shadow."
Multi-image merge: "The person from image 1 wears the outfit from image 2, placed in the background from image 3. Match the lighting so all elements look photographed together."
The most effective editing prompts are imperative and specific about what to preserve. If you want the subject's face unchanged, say so explicitly. If the background should stay the same, include it. Vague instructions like "make it better" produce inconsistent results.
Multi-Image Editing: Combining Up to Three Input Images
Qwen-Image-Edit-2509 and later versions accept up to three input images. The output aspect ratio defaults to that of the first image. Write prompts that reference each image's role explicitly. If you only need to edit one image, right-click the unused input nodes and select Bypass to disable them.
Using ControlNet with Qwen Image Edit
ControlNet adds structural precision without separate preprocessing tools. The model supports depth maps, edge maps, and keypoint maps as control conditions. You need the Union ControlNet LoRA (qwen_image_union_diffsynth_lora.safetensors), which handles all control types in a single file. In the QwenImageDiffsynthControlnet node, set strength values between 0.5 and 0.8 to preserve generative quality while applying the structural constraint.
Expanding What Qwen Image Edit Can Do
Using Qwen-Image LoRA Models for Style and Realism
The Qwen Image community has built a growing library of LoRAs for style transfer, realism, and character consistency. Native LoRA support is built-in, so community LoRAs work without additional calibration nodes.
To apply a LoRA in ComfyUI, add a LoRA Loader node between the model loader and the KSampler. Use a strength of 0.6–1.0 for style LoRAs, and drop it to 0.3–0.5 when stacking multiple LoRAs to avoid compounding artifacts.
Using the Lightning LoRA for Faster Generation
The Lightning LoRA supports 4-step and 8-step inference and runs on as little as 8GB of VRAM with 16GB of system RAM. Quality degradation at 4 steps is minimal, making it worth enabling for rapid iteration.
When using the Lightning LoRA, set CFG to 1.0 and steps to 4. Leaving them at standard values with the LoRA loaded produces blocky, over-saturated output.
Sage Attention for Speed Optimization
For faster generation, consider installing Sage Attention. It reduces attention computation time and meaningfully speeds up generation on compatible hardware. If not installed, disable the Sage Attention node in the workflow. It runs normally without it.
Running Qwen Image from the Command Line
The Qwen CLI handles model download, one-shot queries, interactive chat, and launching a local API-compatible server. Key flags include --prompt, --negative-prompt, --steps, --fast (8-step Lightning LoRA), --ultra-fast (4-step Lightning LoRA v2.0), --seed, --num-images, and --lora. It is especially useful for headless servers or scripted batch jobs where a browser UI is not practical.
Supported Qwen Image Resolutions
Qwen-Image performs well from 768x768 to 2048x2048. Processing time increases significantly at larger sizes. Sticking to native training aspect ratios gives the most reliable results.
| Aspect Ratio | Resolution | Use Case |
|---|---|---|
| 1:1 | 1024x1024 | Default, square format |
| 1:1 | 1328x1328 | Native training resolution |
| 16:9 | 1664x928 | Widescreen, landscape |
| 9:16 | 928x1664 | Mobile, story format |
| 4:3 | 1472x1140 | Landscape photography |
| 3:4 | 1140x1472 | Portrait photography |
In editing mode, Qwen Image Edit outputs match the aspect ratio of the first input image at approximately 1 megapixel. A 768x960 input (4:5 ratio) produces an 896x1160 output, for example. You can override this by connecting a custom latent size node to the workflow.
Last Thoughts on Qwen Image
Qwen-Image leads the open-source field on multi-image composition, instruction-based editing, and bilingual text rendering. The GGUF and FP8 variants bring it within reach of 8 to 16GB GPUs, while the RTX A6000's 48GB covers full BF16 precision without compromise.
See how Qwen-Image compares to other open-source image generation models.
FAQ
What is Qwen Image?
Qwen-Image is a 20B parameter MMDiT (Multimodal Diffusion Transformer) model open-sourced by Alibaba under the Apache 2.0 license. It is the first image generation foundation model from Alibaba's Qwen team, with strong multilingual text rendering and precise image editing capabilities.
What are the Qwen Image Edit VRAM requirements?
BF16 precision requires at least 40GB of VRAM. FP8 requires 16GB. GGUF Q4_K_M requires approximately 14GB. The Q2_K GGUF variant runs on 8GB with lower quality.
How does Qwen Image Edit compare to Flux Kontext?
Qwen Image Edit excels at instruction-based editing, multi-image composition, and bilingual text rendering in English and Chinese. Flux Kontext leads on generation speed. For tasks requiring precise object-level edits, text changes in images, or combining elements from multiple source photos, Qwen Image Edit has a clear advantage.
What is the difference between Qwen Image Edit 2509 and 2511?
The 2511 update (December 2025) improved multi-image editing consistency, added multi-angle view generation from a single reference image, and refined face identity preservation across pose and style transformations. The ComfyUI workflow structure is largely the same, with updated checkpoint files.
Can Qwen Image Edit run on 8GB VRAM?
Yes, using the Q2_K or Q3_K_S GGUF quantization with the Lightning LoRA enabled. Output is usable but noticeably softer than higher-precision variants. At 16GB, FP8 is the sweet spot for near-full quality without heavy quantization.