Qwen Image Edit is one of the most capable open-weights image models. It combines 20 billion parameters with native ControlNet support, multi-image editing, and excellent text rendering.
This guide covers everything from understanding the model to troubleshooting common errors.
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. It is the first image generation foundation model from Alibaba's Qwen team, with significant advances in complex text rendering and precise image editing across multiple languages.
Multimodal Diffusion Transformer (MMDiT): a generative neural network architecture that combines a transformer with a diffusion (denoising) process to process and align multiple data in a single framework.
The architecture combines two large components: The 20.4B DiT Transformer constructs the image The 8.3B Qwen2.5-VL Text Encoder processes the prompt
Qwen-Image's biggest highlight is its breakthrough in complex text rendering. It generates high-quality images with multi-line layouts and paragraph-level content in Chinese and English.
Qwen Image Edit vs. Other AI Image Editors: How Does It Compare?
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 in English and Chinese.
Where Flux leads on generation speed, Qwen Image Edit stands out for instruction-based editing, multi-image composition, and text rendering accuracy.
Unlike cloud solutions that lock you into API pricing, Qwen-Image-Edit-2509 allows editing in a single model that runs entirely on local hardware.
What Is Qwen Image Edit 2509?
Key Improvements Over the Original Qwen Image Edit
Qwen-Image-Edit-2509 is Alibaba's September 2025 monthly iteration of Qwen-Image-Edit. The community has called it a "game changer," primarily for its multi-image editing capabilities and improved editing consistency.
Three areas saw major improvements:
- Person editing: better facial identity preservation across portrait styles and pose transformations.
- Product editing: identity retained when used for product posters.
- Text editing: now supports font type, color, and material changes.
Multi-Image Editing, ControlNet Support, and Advanced Text Editing
Qwen-Image-Edit-2509 adds support for editing multiple images simultaneously, covering "person+person", "person+product", and "person+scene" combinations. It also adds native ControlNet support with built-in depth maps, edge maps, and keypoint maps.
Text editing can now be seamlessly combined with image editing, for example in poster editing scenarios.
Qwen-Image VRAM Requirements: What GPU Do You Need?
VRAM Requirements by Precision: FP16, FP8, and GGUF
Hardware requirements depend heavily on the precision format you choose: BF16 requires at least 40 GB VRAM; FP8 requires 16 GB.
For low-VRAM users, the GGUF format loads into the ComfyUI/models/unet folder and scales down to fit a range of hardware budgets.
| Format | File Size | Min VRAM | Notes |
|---|---|---|---|
| BF16 | 41 GB | 40+ GB | Full precision, best quality |
| FP8 | 21 GB | 16 GB | Good quality, ideal for 24GB GPUs |
| Q8_0 GGUF | 22 GB | 24 GB | Near-original quality |
| Q4_K_M GGUF | 13 GB | 14 GB | Excellent quality, widely recommended |
| Q3_K_M GGUF | 10 GB | 10 GB | Good quality, budget GPUs |
| Q2_K GGUF | 7 GB | 8 GB | Lower quality, extreme constraints |
Running Qwen Image Edit on 8GB, 16GB, and 24GB GPUs
On an 8 GB GPU, 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 16 GB, the FP8 model is the sweet spot for near-full quality without GGUF quantization. The unquantized model needs 24 GB or more (e.g., RTX 4090) and 64 GB of system RAM for optimal performance.
Low-VRAM Options: Quantization and CPU Offloading
Two strategies help when you keep hitting memory limits. Turning on --lowvram in ComfyUI offloads model layers to system RAM between passes, trading speed to reduce VRAM requirements. CPU offloading works similarly by moving DiT layers between passes, enabling inference on limited VRAM.
The 4-step Lightning LoRA reduces sampling steps from 20 down to 4, cutting peak memory use during the diffusion pass. It runs on as little as 8 GB VRAM with 16 GB of system RAM.
To overcome local hardware limitations, Thunder Compute offers RTX A6000 GPUs with 48 GB of VRAM at $0.35/hr. That is enough to run the BF16 model at full precision with no quantization.
Supported Qwen Image Resolutions and Aspect Ratios
Qwen-Image performs well across resolutions from 768×768 to 2048×2048. Processing time increases significantly at larger sizes, affecting speed, text clarity and fine textures.
Sticking to native training aspect ratios gives the most reliable results.
| Aspect Ratio | Resolution | Use Case |
|---|---|---|
| 1:1 | 1024×1024 | Default |
| 1:1 | 1328×1328 | Native training resolution (1:1) |
| 16:9 | 1664×928 | Widescreen |
| 9:16 | 928×1664 | Mobile/story format |
| 4:3 | 1472×1140 | Landscape photography |
| 3:4 | 1140×1472 | Portrait photography |
For editing workflows, the output inherits the aspect ratio of the first input image at approximately 1 megapixel total. You can override this by connecting a custom latent size node
In editing mode, Qwen Image Edit outputs match the aspect ratio of the source image at roughly 1 megapixel. For example, a 768×960 input (4:5 ratio) produces an 896×1160 output.
How to Install Qwen Image Edit in ComfyUI
Make sure your ComfyUI installation is up to date before downloading anything. The nodes and samplers required for Qwen Image Edit were added in recent versions, and an outdated install is a common source of errors. See our ComfyUI usage guide if you need help setting up or updating from scratch.
Downloading the Model Files: Diffusion Model, Text Encoder, and VAE
You need three core files, all available for free on Hugging Face:
qwen_2.5_vl_7b_fp8_scaled.safetensors→ComfyUI/models/text_encoders/qwen_image_vae.safetensors→ComfyUI/models/vae/- Diffusion model (choose one):
qwen_image_edit_2509_fp8_e4m3fn.safetensorsorqwen_image_edit_2509_bf16.safetensors→ComfyUI/models/diffusion_models/
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 under "Custom Nodes Manager."
Setting Up Your ComfyUI 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 a common setup mistake and will prevent models from appearing in their loaders. After placing all files, do a full ComfyUI restart.
Loading the Official Qwen Image Edit 2509 Workflow
Drag the workflow image directly into ComfyUI to load it. ComfyUI decodes the JSON embedded in the PNG metadata and restores the full node graph. Make sure the text encoder and VAE files are already downloaded before doing this.
How to Use Qwen Image Edit 2509 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 like blurriness or watermarks.
Recommended KSampler settings depend on the model variant you are using:
- BF16: CFG 4, 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.
Multi-Image Editing: Combining Up to Three Input Images
Qwen-Image-Edit-2509 accepts up to three input images. The output aspect ratio defaults to that of the first image. Enter your positive prompt at the top and negative prompt at the bottom.
Write prompts that reference each image's role explicitly. A prompt like "The person from image 1 wears the outfit from image 2, placed in the background from image 3" gives the model clear compositional guidance. 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 in ComfyUI
ControlNet adds structural precision without requiring 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. Set the edit instruction in the CLIP Text Encoder, then adjust the strength in the QwenImageDiffsynthControlnet node. Strength values between 0.5 and 0.8 preserve generative quality while still 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.
Speeding Up Generation with the Lightning LoRA
The Lightning LoRA supports 4-step and 8-step inference and runs on as little as 8 GB VRAM with 16 GB of system RAM. Quality degradation at 4 steps is minimal, making it worth enabling for faster iteration.
When using the Lightning LoRA, set CFG to 1.0 and steps to 4. Leaving them at standard values produces blocky, over-saturated output, because the distilled model was trained differently from the base.
Running Qwen Image from the Command Line with Qwen CLI
The Qwen CLI is a lightweight alternative to ComfyUI for terminal-based or automated workflows. It handles model download, one-shot queries, interactive chat, and launching a local API-compatible server from a consistent command interface.
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 (accepts a Hugging Face repo ID). It is especially useful for headless servers or scripted batch jobs where a browser UI is not practical.
Last Thoughts on Qwen Image
Qwen-Image is a massive leap forward for open-weights generative AI. Through this model, Alibaba unlocked a level of instruction-based editing, multi-image composition, and precise multilingual text rendering that was previously restricted to closed-source APIs.
Despite high hardware requirements for full-precision, the ecosystem lets creators can find a setup that fits their local constraints. Qwen Image is a creative engine, whether for generating marketing layouts, orchestrating complex multi-layered photo edits, or automating workflows from the command line.
If you consistently hit VRAM limits locally, Thunder Compute offers cost effective cloud GPUs starting at $0.35/hr. On top of that, you can start instances with ready-to-launch templates for ComfyUI and Forge Neo.