Run Qwen-Image on a cloud GPU

The Qwen image model that fits one 48 GB card is Qwen-Image-2.1. Its files total 33.12 GB, which leaves about 11 GB on the card (calculated), and one checkpoint both generates and edits images. The older Qwen-Image and Qwen-Image-Edit models have 20B parameters and total 57.70 GB, so they need an 80 GB card or an fp8 file. The licenses split the same way: Qwen-Image-2.1 and its Turbo version come under the Qwen Research License, which grants rights for non-commercial purposes only, while Qwen-Image and Qwen-Image-Edit-2511 are Apache 2.0.

Which Qwen image model is which

The table was checked against the Qwen organization on Hugging Face and Qwen's GitHub repositories on 12 November 2026. File totals are the summed safetensors files in each repository, and GB is 10^9 bytes.

Qwen-Image-2.1 edits natively, and the organization has no separate 2.1 editing checkpoint. The Qwen-Image-2.1 card describes a 7B visual generation component with 32 single-stream DiT layers, and the GitHub repository names Qwen3-VL 8B as the text encoder. Two optional prompt rewriters, Qwen-Image-2.1-PE-T2I and Qwen-Image-2.1-PE-I2I, are fine-tuned Qwen3.5-VL 9B checkpoints of 18.82 GB each.

The same repository lists Qwen-Image-2.1 Pro and Turbo as hosted APIs on Alibaba Cloud Model Studio, and the organization has no Pro weights. The Qwen-Image repository lists Qwen-Image-2.0 as launched on 10 February 2026 and points to Qwen Chat, and the organization has no 2.0 weights, so 2.1 is the open 2.x release.

Qwen-Image sizes and memory: 33 GB for 2.1, 58 GB for the 20B models

The left-over rows are calculated as 92% of the card's memory minus the files, so they leave out activations. Neither the Qwen-Image-2.1 card nor the Turbo card gives a memory figure. The GitHub README says only that GPUs with limited memory can use pipe.enable_model_cpu_offload().

The measured numbers come from SGLang's Qwen-Image 2.1 cookbook, which lists peaks measured at 1024 by 1024 pixels and 40 steps: 38.4 GiB on an H200 and on an RTX PRO 6000, and 22.7 GiB on a 24 GB RTX 4090 with the encoder offloaded layer by layer. 38.4 GiB is about 41.2 GB, inside the 44.2 GB that a 48 GB card has after the 8% margin (calculated). The 38.4 GiB figures come from 96 and 141 GB cards, and the page gives no peak for the native 2048 by 2048 size or for edits with several references. For the 20B models, ComfyUI's Qwen-Image tutorial shows its fp8 workflow using 86% of the memory of a 24 GB RTX 4090D.

Which GPU to rent for Qwen-Image

My pick for Qwen-Image-2.1 at about 1 megapixel is a 48 GB card. For 2K output, edits with several reference images, or the 20B models in BF16, I would take the RTX PRO 6000, because its 96 GB leaves about 55 GB after the 2.1 files and about 31 GB after the 20B files (calculated), room that no published figure measures. An 80 GB card holds the 20B files with about 16 GB left. The 48 GB page and the RTX PRO 6000 page compare the cards.

Qwen-Image and Qwen-Image-Edit in ComfyUI

ComfyUI's release notes list Qwen-Image 2.1 support in v0.37.0, published on 21 September 2026. The ComfyUI template is built from an image that DeployGPU rebuilds, so check the version in the app before you download anything. The models guide shows how.

ComfyUI's Qwen-Image-2.1 tutorial lists text-to-image, image edit and background removal workflows, plus Turbo versions of the first two, in the Template Library under "Qwen-Image-2.1" and "Qwen Image 2.1 Turbo". The base workflows sample 25 steps at cfg 1 and the Turbo ones 8 steps. They load the int8 files below from the Comfy-Org/Qwen-Image-2.1 repository, which is not gated.

The base set totals 17.28 GB (calculated), so a 48 GB card has about 27 GB left. The tutorial calls the BF16 files full precision that needs more memory: qwen_image_2.1_bf16.safetensors is 14.23 GB and qwen3vl_8b_bf16.safetensors is 17.53 GB. Inside the container, the folders sit under /workspace/runpod-slim/ComfyUI/models/, and sudo docker exec deploygpu-workload ls /workspace/runpod-slim/ComfyUI/models confirms the layout. The block below opens a shell in the container and downloads from inside it:

sudo docker exec -it deploygpu-workload bash
cd /workspace/runpod-slim/ComfyUI/models
curl -L --fail -o diffusion_models/qwen_image_2.1_int8_convrot.safetensors \
  https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/diffusion_models/qwen_image_2.1_int8_convrot.safetensors
curl -L --fail -o text_encoders/qwen3vl_8b_int8_convrot.safetensors \
  https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/text_encoders/qwen3vl_8b_int8_convrot.safetensors
curl -L --fail -o vae/qwen_image_2.1_vae_bf16.safetensors \
  https://huggingface.co/Comfy-Org/Qwen-Image-2.1/resolve/main/vae/qwen_image_2.1_vae_bf16.safetensors

For Qwen-Image-Edit-2511, ComfyUI's tutorial loads qwen_image_edit_2511_bf16.safetensors at 40.86 GB, qwen_2.5_vl_7b_fp8_scaled.safetensors at 9.38 GB and qwen_image_vae.safetensors at 0.25 GB. That is 50.50 GB (calculated), more than a 48 GB card's 44.2 GB. The Comfy-Org/Qwen-Image-Edit repository also has qwen_image_edit_2511_fp8mixed.safetensors at 20.53 GB, which brings the set to 30.17 GB. The template installs no models and stopping deletes the disk, so the template page and the stop guide matter for repeat sessions.

Qwen-Image-2.1 with Diffusers

The Turbo card installs Diffusers from source with Transformers 5.17 or newer. The release notes of Diffusers 0.41.0, published on 6 October 2026, list the Qwen-Image 2.1 pipeline and the sampling-sigma change that Turbo needs. Install a CUDA build of PyTorch first, as the PyTorch install guide explains.

pip install git+https://github.com/huggingface/diffusers.git
pip install "transformers>=5.17.0" accelerate pillow

This is the code from the Qwen-Image-2.1 card, with its text-to-image and editing snippets joined so they share one pipeline:

import torch
from PIL import Image
from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained(
    "Qwen/Qwen-Image-2.1", torch_dtype=torch.bfloat16
).to("cuda")

image = pipe(
    prompt="A neon shop sign that reads \"QWEN IMAGE 2.1\", rainy night, reflections on wet pavement",
    width=2048, height=2048,
    num_inference_steps=40,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("t2i_example.png")

input_image = Image.open("input.png")
image = pipe(
    prompt="Change the background to a sunset beach",
    image=input_image,
    num_inference_steps=40,
    generator=torch.Generator("cuda").manual_seed(42),
).images[0]
image.save("edit_example.png")

The Turbo card loads Qwen/Qwen-Image-2.1-Turbo with the same class. Its checkpoint saves an 8-step schedule, and the card says setting num_inference_steps alone does not override it. The Qwen-Image-Edit-2511 card uses QwenImageEditPlusPipeline with a list of images and 40 steps. AI Toolkit's README lists Qwen-Image, Qwen-Image-2512, Qwen-Image-2.1 and 2.1 Turbo as supported base models, and the LoRA guide covers its install on a plain GPU VM.

The license of each Qwen image model

The Qwen Research License grants the right "to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials FOR NON-COMMERCIAL PURPOSES ONLY" in section 2.a. Section 1.i defines "Non-Commercial" as "for research or evaluation purposes only". Section 2.b says "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us", and the license names an address for requesting one. Section 4.b says that if you use the Materials or any outputs or results from them to create, train, fine-tune, or improve an AI model that is distributed or made available, you shall prominently display "Built with Qwen" or "Improved using Qwen" in the related product documentation. Section 8 makes the laws of China govern and gives the People's Courts in Hangzhou City exclusive jurisdiction.

Qwen-Image carries the Apache License 2.0 as its LICENSE file, and its card says "Qwen-Image is licensed under Apache 2.0." The Edit, Edit-2509 and Edit-2511 cards repeat that sentence, and the Qwen-Image-2512 card has only the apache-2.0 metadata. The GitHub repository for Qwen-Image-2.1 says it is under the Qwen Research License Agreement, and the Qwen-Image repository says Apache 2.0. For other image models on the same template, the FLUX page covers FLUX.3 [klein] and the Z-Image page covers a 6B model that fits much smaller cards.