How to run Qwen-Image-Edit in ComfyUI on a cloud GPU
Rent a GPU on Deploy, then open the ComfyUI template or install ComfyUI from the official repository. Follow the ComfyUI Qwen-Image-Edit tutorials: put each weight file in the folder that page prints, load the image and the prompt, queue the workflow, and copy the output folder off the instance before you destroy it. Destroying the instance deletes its disk.
The diffusers quick start for the same weights, including the current Hub id, is How to run Qwen-Image on a cloud GPU. Step1X-Edit’s README prints peak GPU memory from 18GB to 49.8GB and recommends 80GB: How to run Step1X-Edit on a cloud GPU. FLUX.1 Kontext in ComfyUI is How to run FLUX.1 Kontext in ComfyUI on a cloud GPU. The ComfyUI install, including port 8188, is How to install ComfyUI and ComfyUI-Manager on a cloud GPU. If the process runs out of memory: How to fix CUDA out of memory on a cloud GPU. How published FLUX and SDXL weight sizes are read: How much VRAM do I need for FLUX and SDXL.
The model cards for Qwen/Qwen-Image-Edit, Qwen/Qwen-Image-Edit-2509, and Qwen/Qwen-Image-Edit-2511 say Qwen-Image is licensed under Apache 2.0 (read 2026-11-13). Each card’s front matter is license: apache-2.0. The model API that day reported gated false, top-level license null, and cardData.license apache-2.0 for all three. lastModified was 2025-08-25T04:41:11.000Z, 2025-09-22T13:37:12.000Z, and 2025-12-23T13:13:52.000Z. The Qwen/Qwen-Image-Edit card says the editor is built upon the 20B Qwen-Image model. That parameter count is not a VRAM figure. Read the license before you redistribute the weights.
The steps below are taken from the ComfyUI pages named in each section. They were not run on a CloudGPU instance for this page. Anything this page adds around them is marked.
What the tutorials and the model cards say about memory
The Qwen-Image-Edit tutorial and the Qwen-Image-Edit-2511 tutorial, both read 2026-11-13, publish no GPU GB figure. The three model cards above also publish no GPU GB figure. The edit tutorial calls Qwen-Image-Edit a 20B model and a 20B MMDiT in the workflow heading. Those are parameter counts. The text-to-image tutorial’s file sizes and its RTX 4090D usage table are a different page. How to run Qwen-Image on a cloud GPU leaves that table on the tutorial. This page does not copy it, and it does not treat those file sizes as a VRAM requirement for the edit workflows.
GET /v1/public/compute-capabilities does not list GPUs. It returns feature flags. The memory column is vram_gb from gpu-prices.json, updated_at 2026-11-13T17:08:25.407Z. GET /v1/public/gpu-catalog the same day (generated_at 2026-11-13T17:08:26.374Z) listed the same names and the same vram_gb values, plus Ascend 910B at 64, which is not in gpu-prices.json and is not listed here. The catalog response also includes an availability field. This page does not copy it. Every prices row has gpu_count 1. Every catalog row has min_gpu_count 1. There is no upstream GPU GB cutoff on the edit tutorials to match these values against. Which listing you can start is on Deploy.
The catalog template list included ComfyUI. It had no Qwen-Image-Edit template. Whether that ComfyUI template already contains these files was not checked. How to choose a rental GPU for ComfyUI is the card page for that template.
| Catalog name | vram_gb |
|---|---|
| RTX 2080 | 8 |
| RTX 3080 | 10 |
| RTX 2080 Ti | 11 |
| RTX 3080 Ti | 12 |
| V100 16G | 16 |
| A30 | 24 |
| RTX 3090 | 24 |
| RTX 4090 | 24 |
| RTX 4090D | 24 |
| RTX 4080 SUPER 32G | 32 |
| RTX 5090 | 32 |
| RTX 5090D | 32 |
| V100 32G | 32 |
| A100 40G | 40 |
| A800 40G | 40 |
| A40 | 48 |
| RTX 4090 48G | 48 |
| RTX 4090D 48G | 48 |
| RTX A6000 | 48 |
| NVIDIA RTX PRO 5000 72G | 72 |
| RTX PRO 5000 72G | 72 |
| A100 80G | 80 |
| A800 80G | 80 |
| H800 | 80 |
| RTX 6000D | 83 |
| H20 96G | 96 |
| RTX PRO 6000 96G | 96 |
| RTX PRO 6000 Max-Q | 96 |
| H20-3e 141G | 141 |
Qwen-Image-Edit files and workflow
On Deploy, choose the ComfyUI template, or follow How to install ComfyUI and ComfyUI-Manager on a cloud GPU on a Raw GPU. The edit tutorial says to update ComfyUI, then download the workflow JSON or search “Qwen Image Edit” in the template library. It also says that if the template is missing, ComfyUI may be outdated. A sentence on that page says cloud will update after the ComfyUI stable release. That sentence is about the note on the ComfyUI docs page. This page did not check the CloudGPU template’s ComfyUI version.
The same page says all models can be found at Comfy-Org/Qwen-Image_ComfyUI or Comfy-Org/Qwen-Image-Edit_ComfyUI. On 2026-11-13 those two Hub URLs returned HTTP 307 to Comfy-Org/Qwen-Image and Comfy-Org/Qwen-Image-Edit. The model API for the redirect targets that day reported gated false, top-level license null, and cardData.license apache-2.0. lastModified was 2026-09-23T08:06:14.000Z and 2026-09-23T08:05:05.000Z. The file list of Comfy-Org/Qwen-Image-Edit included split_files/diffusion_models/qwen_image_edit_fp8_e4m3fn.safetensors. The file list of Comfy-Org/Qwen-Image included split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors and split_files/vae/qwen_image_vae.safetensors. The tutorial’s own tree is not the split_files/ layout. Copying the Hub layout into ComfyUI/models would miss the tree the tutorial shows. Place each file in the folder the tutorial prints. The tutorials do not print an hf download command. The Hugging Face download guide is the hf download reference. This page does not add a download command.
The edit tutorial’s storage tree, copied:
ComfyUI/
├── models/
│ ├── diffusion_models/
│ │ └── qwen_image_edit_fp8_e4m3fn.safetensors
│ ├── loras/
│ │ └── Qwen-Image-Lightning-4steps-V1.0.safetensors
│ ├── vae/
│ │ └── qwen_image_vae.safetensors
│ └── text_encoders/
│ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
The diffusion file link on that page is qwen_image_edit_fp8_e4m3fn.safetensors under Comfy-Org/Qwen-Image-Edit_ComfyUI. The text encoder and VAE links are qwen_2.5_vl_7b_fp8_scaled.safetensors and qwen_image_vae.safetensors under Comfy-Org/Qwen-Image_ComfyUI. The LoRA link is Qwen-Image-Lightning-4steps-V1.0.safetensors under lightx2v/Qwen-Image-Lightning.
The tutorial’s steps: the Load Diffusion Model node loads qwen_image_edit_fp8_e4m3fn.safetensors, the Load CLIP node loads qwen_2.5_vl_7b_fp8_scaled.safetensors, and the Load VAE node loads qwen_image_vae.safetensors. The Load Image node takes the image to edit. The prompt is set in the CLIP Text Encoder node. The Scale Image to Total Pixels node scales the input to a total of one million pixels. The page says that is mainly to avoid quality loss from oversized inputs such as 2048x2048, and that Ctrl+B bypasses that node. The same page says that to use the “4-step Lighting LoRA”, select the LoraLoaderModelOnly node and press Ctrl+B to enable it. The filename in the tree is Qwen-Image-Lightning-4steps-V1.0.safetensors. The spelling “Lighting” is the sentence on the tutorial. The page says the Ksampler steps and cfg have a note under the node. This page does not copy a number the note does not print in the text. Then click Queue, or use Ctrl (cmd) + Enter. The input filename the page shows for the Load Image node is image_qwen_image_edit_input_image.png.
Qwen-Image-Edit-2509
A later section of the same tutorial says Qwen-Image-Edit-2509 keeps the text editing of the original model and adds multi-image input, more consistent edits across a series of images, and ControlNet support. Search “Qwen Image Edit 2509” in the template library, or download that section’s workflow JSON. Its storage tree, copied:
ComfyUI/
├── models/
│ ├── diffusion_models/
│ │ └── qwen_image_edit_2509_fp8_e4m3fn.safetensors
│ ├── loras/
│ │ └── Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors
│ ├── vae/
│ │ └── qwen_image_vae.safetensors
│ └── text_encoders/
│ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
The diffusion link is qwen_image_edit_2509_fp8_e4m3fn.safetensors under Comfy-Org/Qwen-Image-Edit_ComfyUI. The LoRA line says it is optional, for 4-step Lightning acceleration, and the link is Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors under lightx2v/Qwen-Image-Lightning. The text encoder and VAE are the same two files as the section above. The input filename that section shows is image_qwen_image_edit_2509_input_image.png.
The ComfyUI examples file qwen_image/README.md, examples HEAD f9431bb000ce792094ff345446e22cac1ea6cef3 committed 2025-11-26T20:33:58Z, also says to put qwen_image_edit_fp8_e4m3fn.safetensors and qwen_image_edit_2509_fp8_e4m3fn.safetensors in ComfyUI/models/diffusion_models, and to use the text encoder and VAE from its basic workflow. That README does not name Qwen-Image-Edit-2511. edit_models/README.md in the same checkout does not name Qwen-Image-Edit. Dragging the example image is the workflow step that README prints.
Qwen-Image-Edit-2511
The edit tutorial’s workflow sections do not name Qwen-Image-Edit-2511. A separate page does: Qwen-Image-Edit-2511 in ComfyUI, read the same day. It says 2511 is an enhanced version of Qwen-Image-Edit-2509, and it describes consistency, multi-person edits, integrated LoRAs, and material replacement. It publishes no parameter count and no GPU GB figure. Search “Qwen-Image-Edit-2511” in the template library, or drag the workflow preview. Its material-replacement storage tree, copied:
ComfyUI/
├── models/
│ ├── text_encoders/
│ │ └── qwen_2.5_vl_7b_fp8_scaled.safetensors
│ ├── loras/
│ │ └── Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors
│ ├── diffusion_models/
│ │ └── qwen_image_edit_2511_bf16.safetensors
│ └── vae/
│ └── qwen_image_vae.safetensors
The diffusion link on that page is qwen_image_edit_2511_bf16.safetensors under Comfy-Org/Qwen-Image-Edit_ComfyUI. The LoRA link is Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors under lightx2v/Qwen-Image-Edit-2511-Lightning, and the page marks that LoRA optional. The file list of Comfy-Org/Qwen-Image-Edit on 2026-11-13 included split_files/diffusion_models/qwen_image_edit_2511_bf16.safetensors and split_files/diffusion_models/qwen_image_edit_2511_int8_convrot.safetensors. The input filenames that section shows are leather_sofa.png and texture_fur.png.
A second section on the same page, “Qwen-Image-Edit-2511 Multiple Angles with the 3D Camera Node”, says to search “Generate Multiple Angles with 3D Camera Node”. It says turbo mode (enable_turbo_mode) is on by default: 4 steps, CFG 1, and the 4-step Lightning LoRA. It says switching turbo mode off is the 40-step, CFG 4 pass without the Lightning LoRA. Those two pairs are the page’s. This page did not remeasure them. The diffusion file in that section’s tree is qwen_image_edit_2511_int8_convrot.safetensors, placed in ComfyUI/models/diffusion_models/. The text encoder and VAE filenames match the tree above. The multiple-angles file the page names is qwen-image-edit-2511-multiple-angles-lora.safetensors in ComfyUI/models/loras/. The Hugging Face blob link for that LoRA uses a repository path owned by a hosted-inference vendor. Copy that URL from the tutorial. This page does not print the path. The input filename that section shows is cyber_genesis.png.
The diffusers id Qwen/Qwen-Image-Edit-2511 is not any of the Comfy-Org filenames above. The load call that uses it is on How to run Qwen-Image on a cloud GPU.
Web UI
The two ComfyUI tutorials do not document a Gradio launch. ComfyUI is the UI. comfy/cli_args.py, as the install guide records it, defaults --listen to 127.0.0.1 and --port to 8188. The tunnel is How to use SSH port forwarding on a cloud GPU. The ComfyUI template’s own address is on Deploy. The Qwen-Image repository’s Gradio command is a different route, on the Qwen-Image guide. This page does not copy it.
When it fails
CUDA out of memory. The edit tutorials and the three model cards publish no GPU GB figure and no offload flag. If the process raises that error, use how to fix CUDA out of memory on a cloud GPU.
“CUDA error: no kernel image is available for execution on the device” on an RTX 5090. The edit tutorials do not name a PyTorch build or compute capability 12.0. The RTX 5090 guide says that card is compute capability 12.0 (sm_120) and that a PyTorch build for CUDA 12.8 or newer is the line that announced support. The ComfyUI install guide records the README’s cu130 note for Nvidia 20 series and above. This page did not install a torch wheel and did not re-test this workflow on an RTX 5090. Catalog RTX 5090 and RTX 5090D are 32. That is the catalog number only.
Save the images, then destroy
The edit tutorials do not name the directory ComfyUI writes at runtime. The sample result linked from the first tutorial is output/image_qwen_image_edit.jpg in the Comfy-Org/workflow_templates repository. That path is the docs sample, not a path this page created. The install guide records that folder_paths.py sets output_directory to os.path.join(base_path, "output"). Copy that directory, and any workflow you exported, before you destroy the instance. Save your work before you destroy a cloud GPU is the checklist. The terms say data on local instance storage is permanently deleted when the instance is destroyed.
The pricing page FAQ states the billing rule in these words:
Per-minute options have a 1-minute minimum; unused prepaid minutes are refunded when you destroy the instance. Some configurations use whole-hour billing with a 1-hour minimum; unused time within a charged hour is not refunded on voluntary stop. The billing period, minimum charge and initial prepayment are shown before confirmation. No subscription.
Prices are not copied here.
Frequently asked questions
How much VRAM do the Qwen-Image-Edit tutorials print?
The Qwen-Image-Edit tutorial and the Qwen-Image-Edit-2511 tutorial publish no GPU GB figure (read 2026-11-13). The model cards for Qwen/Qwen-Image-Edit, Qwen/Qwen-Image-Edit-2509, and Qwen/Qwen-Image-Edit-2511 also publish no GPU GB figure. The edit tutorial calls the model a 20B MMDiT. That count is not a VRAM figure. This page does not mark a catalog row as enough. Which listing you can start is on https://cloudgpu.app/deploy.
Where does each weight file go?
The Qwen-Image-Edit tree places qwen_image_edit_fp8_e4m3fn.safetensors in models/diffusion_models, Qwen-Image-Lightning-4steps-V1.0.safetensors in models/loras, qwen_image_vae.safetensors in models/vae, and qwen_2.5_vl_7b_fp8_scaled.safetensors in models/text_encoders. The 2509 section swaps in qwen_image_edit_2509_fp8_e4m3fn.safetensors and Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16.safetensors. The 2511 material-replacement tree uses qwen_image_edit_2511_bf16.safetensors and Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors. The multiple-angles section uses qwen_image_edit_2511_int8_convrot.safetensors and qwen-image-edit-2511-multiple-angles-lora.safetensors. The Hub copies of the diffusion, text-encoder, and VAE files sit under split_files/. The tutorial trees do not.
Where does ComfyUI save the edited image?
The tutorials do not name the runtime output directory. The install guide records that folder_paths.py sets output_directory to the output folder under the ComfyUI base path. The file output/image_qwen_image_edit.jpg linked from the first tutorial is a sample in Comfy-Org/workflow_templates. Copy the output directory before you destroy the instance. This page did not run the workflow.
What license do the Qwen-Image-Edit model cards state?
Each of Qwen/Qwen-Image-Edit, Qwen/Qwen-Image-Edit-2509, and Qwen/Qwen-Image-Edit-2511 says Qwen-Image is licensed under Apache 2.0. The front matter is license: apache-2.0. The model API on 2026-11-13 reported gated false and cardData.license apache-2.0. The license section this page read prints no dollar amount.
Sources (read 2026-11-13, Asia/Shanghai)
- ComfyUI Qwen-Image-Edit tutorial (filenames, folders, one million pixels, 2048x2048, the “4-step Lighting LoRA” sentence, Queue; no GPU GB figure): https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit
- ComfyUI Qwen-Image-Edit-2511 tutorial (
qwen_image_edit_2511_bf16.safetensors,qwen_image_edit_2511_int8_convrot.safetensors, turbo mode 4 steps CFG 1, 40-step CFG 4; no GPU GB figure): https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit-2511 - ComfyUI examples
qwen_image/README.md(edit and edit-2509 filenames; 2511 absent) andedit_models/README.md(does not name Qwen-Image-Edit): examples HEADf9431bb000ce792094ff345446e22cac1ea6cef3, committed 2025-11-26T20:33:58Z - Model cards
Qwen/Qwen-Image-Edit,Qwen/Qwen-Image-Edit-2509, andQwen/Qwen-Image-Edit-2511(Apache 2.0, 20B sentence on the first card, no GPU GB figure): read 2026-11-13. Model API the same day:gatedfalse,cardData.licenseapache-2.0 - Hub redirects:
Comfy-Org/Qwen-Image-Edit_ComfyUItoComfy-Org/Qwen-Image-Edit, andComfy-Org/Qwen-Image_ComfyUItoComfy-Org/Qwen-Image, both HTTP 307 on 2026-11-13. File lists of the redirect targets included the diffusion, text-encoder, and VAE names cited above - ComfyUI
folder_paths.pyoutput directory andcomfy/cli_args.pyport 8188: recorded on How to install ComfyUI and ComfyUI-Manager on a cloud GPU - Public GPU memory: https://cloudgpu.app/gpu-prices.json (
updated_at2026-11-13T17:08:25.407Z) and https://cloudgpu.app/v1/public/gpu-catalog (generated_at2026-11-13T17:08:26.374Z).GET /v1/public/compute-capabilitiesreturned feature flags and no GPU list. The catalog template list included ComfyUI and had no Qwen-Image-Edit template
Cards you can start are on the Deploy page. Prices are on https://cloudgpu.app/pricing and are not repeated here.
Last updated: 2026-11-13 (Asia/Shanghai).