AI image models have become incredibly capable, but using them effectively often means juggling multiple models. You might need one model for text-to-image generation, another for image editing, another for background removal, and yet another workflow for handling transparent assets.
Qwen-Image-2.1 is an open-source 7B unified text-to-image generation and image editing model from the Qwen family. What makes it interesting is its relatively compact visual generation component, with just 7B parameters built around 32 Single-Stream DiT layers.
 |
| Qwen image 2.1 showcase |
The goal isn't simply to make another image generator. Instead, Qwen-Image-2.1 combines generation and editing into a single model while focusing on inference efficiency, reference-image control, transparency, and visual quality.
 |
| qwen image 2.1 architecture |
Qwen-Image-2.1 uses a lightweight architecture with mixed-granularity attention and prefix KV cache reuse. The idea is to reduce unnecessary computation while maintaining strong image quality. This makes the model particularly interesting for users who want to experiment with image generation and editing without relying on extremely large visual models.
Instead of generating an image with a background and removing it afterward, the model can directly work with transparency. This is useful for workflows involving: Product assets, Character cutouts, Stickers, Icons, Design elements , E-commerce imagery. The model supports up to 10 reference images in a single image edit workflow.
Installation
1. You should have the ComfyUI already installed. Update to its latest version if already have using Manager by selecting Update ComfyUI.
2. Download one of the following Qwen Image 2.1 diffusion models depending on your available VRAM.
(a) qwen_image_2.1_bf16.safetensors - BF16 variant, Higher memory requirement
Used in the VIP workflow and custom workflows
(b) qwen_image_2.1_int8_convrot.safetensors -Quantized INT8 version, requires significantly less VRAM. Save any of them into your ComfyUI/models/diffusion_models folder.
Alternative(For Low VRAMS inference):
Instead of diffusion models, you can also install the Qwen image 2.1 GGUF (Q3-Q8) models. Save it into your ComfyUI/models/unet folder. Make sure you have the ComfyUI-GGUF custom node by city96 already installed.
3. Qwen Image 2.1 uses a Qwen3-VL 8B text encoder. Again, you can choose the version according to your available VRAM.
(a) qwen3vl_8b_bf16.safetensors - Full precision
that needs highest memory requirement
(b) qwen3vl_8b_int8_convrot.safetensors - Quantized INT8 version that runs on lower VRAM
(c) qwen3vl_8b_w4a8.safetensors - lowest memory requirement of the three, intended for GPUs with more limited VRAM
Place the selected text encoder into your ComfyUI/models/text_encoders folder.
4. Download VAE (qwen_image_2.1_vae_bf16.safetensors) is the RGBA autoencoder, uses 16x spatial compression and save it into your ComfyUI/models/vae folder. This VAE is required for both text-to-image and image-editing workflows.
5. Use Prompt rewriting enhancement models but its optional. These models are used by the prompt-rewriting/enhancement nodes. Both are separate models and are not interchangeable, means you need both.
(a) qwen3.5_9b_qwen_image_2.1_pe_t2i.int8_convrot.safetensors -Its Text-to-Image Prompt Enhancer, that rewrites and enhances your text-to-image prompts
(b) qwen3.5_9b_qwen_image_2.1_pe_i2i.int8_convrot.safetensors - Its an Image-Edit Prompt Enhancer, rewrites and enhances prompts for image-editing workflows. Save them into ComfyUI/models/text_encoders folder.
6. Restart your ComfyUI to get the effect.
Workflow
1. Download Qwen Image 2.1 workflows from our Hugging face repository.
(a) Qwen_Image_2.1_Text2Img.json
(b) Qwen_Image_2.1_Edit.json
2. Drag and drop into ComfyUI canvas. Load all the models (diffusion models, text encoders, vae) into their respective nodes.
3. Set configurations with K sampler Settings-
CFG- 1.0
Steps- 25 for normal, 40-50 for best quality output
sampler-Euler
Scheduler-simple
Resolution aspect ratio supported-
1:1 (square)
2:3 (Photo style)
3:2 (portrait)
4:3 (Standard)
9:16 (Portrait widescreen)
16:9 (Widescreen)
21:9 (Ultrawidescreen)
megapixel- 2048x2048 (maximum); default is set to 1024x1024
You can also check our X post for default Txt to Image output, sampler + scheduler output and image edit output.
4. Do the prompting:
For Text to Image - Add positive prompts into prompt box, negative not needed. Both English and Chinese prompting language supported.
Its also support transparent image generation. Use this prompting style for transparent images - This is an RGBA format image with transparency. [your description]. The image has an alpha channel and a transparent background.
For Image Edits- the model supports maximum 10 input images. You can upload single image and do the editing or upload your multiple reference images into the edit workflow. Image 1 is the subject as target and the rest will be the references. Use any of the prompting style examples, or create your own :
(a) replace the texture of wall on <image 1> with the green texture of <image 2>
(b) replace the pose of <image 1> with the pose of <image 2>. keep the character of <image 1>
(c) replace the lighting and background on <image 1> with the lighting and background of <image 2>
4 - 7. transfer the style of <image 1> with the style of <image 2> + more conditionals
(d) girl of <image 1> is sit on the yellow chair of <image 2>
(e) right side view low-angle shot t medium shot of <image 1>
(f) <image 1> at night backlight
(g) Upscale the <image 1>
5. Hit Run to start generation.
Text to Image:
Prompt:
Photorealistic full-body photograph of an adult woman in a small neighborhood convenience store, long wavy auburn-red hair, fair skin, calm slightly serious expression, looking directly at the camera with her head subtly tilted. She stands beside a checkout counter, both arms extended forward, holding a glass bottle of dark amber beverage on the counter. Fitted black off-the-shoulder crop top, high-waisted faded charcoal-gray distressed denim cutoff shorts with frayed hems, white low-top sneakers. Casual contemporary streetwear. Convenience-store interior, checkout counter and cash register on the left, colorful snack shelves behind the counter, tall glass-door refrigerators packed with beverages and groceries on the right, tiled floor, suspended acoustic ceiling with rectangular fluorescent lights. Strong perspective lines from the counter and refrigerator aisle. Vertical full-body composition, subject centered slightly right, 28–35mm lens, realistic proportions, eye-level/slightly low camera angle, natural candid pose, authentic retail lighting, soft shadows, realistic skin texture, detailed hair strands, denim fibers, glass reflections and product packaging, documentary photography, high-resolution DSLR photo, photorealistic, natural colors, subtle imperfections.
Upscaling as Image edit:
 |
| Input image as low quality |
 |
| upscaled image as high quality |
Prompt:
Upscale the <image 1>
Qwen-Image-2.1 is interesting because it isn't trying to solve image generation as an isolated problem. It's moving toward the idea of a unified image workspace where generation, editing, reference conditioning, masking, and transparency can happen within the same model.
The 7B visual generation component is also notable. Instead of relying entirely on scale, the model focuses on architectural efficiency while adding capabilities that are genuinely useful in practical creative workflows.
