Running Qwen-Image-Edit-2511 on RunPod

In the previous article, I checked the specs for running Qwen-Image-Edit-2511 locally: RTX 4090 at ¥450,000, 64GB RAM at over ¥100,000. As of January 2026, building an AI PC is terrible timing.

So I went with a cloud GPU instead (RunPod). RTX 4090 at $0.34/hour (roughly ¥51), with each image generation costing about ¥2–3. For light users, cloud wins by a mile.

What Is Qwen-Image-Edit-2511?

An image editing model with 20B parameters released by Qwen (Tongyi Qianwen). It supports text-instruction-based image editing, inpainting, and outpainting.

What I actually wanted to use is fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA. It’s a LoRA that lets you specify 96 camera angles (4 elevations × 8 azimuths × 3 distances) to regenerate images. Built from 3D model training data, it can transform a character illustration into a view from a different angle.

What Is RunPod?

A cloud GPU service. 5–6× cheaper than AWS or GCP, and it comes with ComfyUI templates so setup is easy.

GPU	Community Cloud	Spot (with interruption risk)
RTX 4090	$0.34/hr (¥51)	~50% off
RTX 3090	$0.22/hr (¥33)	~50% off

What You Need

RunPod account
Credit card or PayPal (charge $10+)
1–2 hours for initial setup

Step 1: Create a RunPod Account

Go to https://www.runpod.io/
Sign up (Google/GitHub login available)
In the left menu: Billing → Add Credits, add about $10

Step 2: Launch a Pod

Left menu: Pods → + Deploy
Select GPU: RTX 4090 (Community Cloud)
Search for “ComfyUI” in the template search and select it
Container Disk: 50GB
Volume Disk: 100GB (for model storage, reusable across sessions)
Click Deploy

Takes 1–2 minutes to start.

Step 3: Access ComfyUI

Click “Connect” on the running Pod → HTTP Service (port 3000) opens ComfyUI.

The default workflow appears on first load. If you see this screen, you’re good.

Step 4: Install Custom Nodes

Click the “Manager” button in the top-right of ComfyUI (if ComfyUI Manager is pre-installed).

If Manager is missing, connect via SSH and install manually.

SSH Connection (only if needed)

Click “Connect” → SSH over exposed TCP to get the connection details.

ssh root@<ip> -p <port> -i ~/.ssh/id_rsa

Installing Custom Nodes

cd /workspace/ComfyUI/custom_nodes

# Camera Angle Selector (96-angle selection UI)
git clone https://github.com/NickPittas/ComfyUI_CameraAngleSelector.git

# ComfyUI Manager (if not installed)
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

Choosing a Model Variant

Qwen-Image-Edit-2511 comes in multiple quantized variants. Choose based on your GPU’s VRAM.

Variant	Size	Required VRAM	Quality	Use Case
BF16 (full precision)	~57GB	40GB+	Best	A100/H100
FP8	~20GB	~20GB	Nearly equivalent	Recommended for RTX 4090
NF4	—	~17GB	Good	RTX 3090
GGUF Q4_K_M	~13GB	12GB+	Practical	RTX 3060

FP8 is optimal for RTX 4090 (24GB). Maintains quality with VRAM headroom.

Recommended Settings by Variant

Variant	CFG	Steps	Notes
BF16	4.0	40	High quality, slow
FP8	4.0	20	Good balance
FP8 + Lightning LoRA	1.0	4	Fast preview

Download Sources

FP8: 1038lab/Qwen-Image-Edit-2511-FP8
GGUF: unsloth/Qwen-Image-Edit-2511-GGUF
Lightning (fast inference): lightx2v/Qwen-Image-Edit-2511-Lightning

The following steps assume FP8. For full precision, replace the download source with Qwen/Qwen-Image-Edit-2511.

Step 5: Download Models

Download the Qwen-Image-Edit-2511 base model and LoRA. Even the FP8 version is 20GB+, so it takes a while.

cd /workspace/ComfyUI/models

# Install HuggingFace CLI
pip install -U huggingface-hub

# Base model FP8 (~20GB)
huggingface-cli download 1038lab/Qwen-Image-Edit-2511-FP8 \
  --local-dir ./checkpoints/Qwen-Image-Edit-2511-FP8/

# Multiple-Angles-LoRA
mkdir -p loras
huggingface-cli download fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA \
  --local-dir ./loras/Qwen-Image-Edit-2511-Multiple-Angles-LoRA/

For full precision:

# Base model BF16 (~57GB)
huggingface-cli download Qwen/Qwen-Image-Edit-2511 \
  --local-dir ./checkpoints/Qwen-Image-Edit-2511/

Lightning LoRA for Fast Inference (Optional)

Reduces inference steps from 40 to 4. Quality drops slightly but speed improves dramatically.

# Lightning LoRA
huggingface-cli download Kijai/flux-fp8 \
  Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16.safetensors \
  --local-dir ./loras/

Step 6: Restart ComfyUI

Restart ComfyUI so it picks up the new models and nodes.

supervisorctl restart comfyui

Or stop and start the Pod itself.

Step 7: Build the Workflow

Connect the following nodes in ComfyUI.

Basic Setup

[Load Checkpoint] Qwen-Image-Edit-2511
       ↓
[Load LoRA] Multiple-Angles-LoRA (strength: 0.8~1.0)
       ↓
[Load Image] Input image
       ↓
[Camera Angle Selector] Select angle
       ↓
[CLIP Text Encode] Prompt
       ↓
[KSampler]
       ↓
[Save Image]

Using Camera Angle Selector

A 3D UI appears where you select from 96 angles.

Direction: front, front-right, right, back-right, back, back-left, left, front-left
Height: overhead, eye level, low angle, ground level
Distance: close-up, medium shot, full shot

The selection outputs a prompt like <sks> front view eye level medium shot.

Recommended Parameters

Parameter	Fast Mode	High Quality Mode
Steps	4 (with Lightning LoRA)	20
CFG Scale	3.0	4.0
Sampler	euler	dpm++_2m_karras

Cost Management

RunPod charges by the hour, so manage your pods when done.

Stop: Pauses GPU billing, Volume is retained but storage billing continues
Delete: Removes Pod, you can choose to keep the Volume
Terminate: Completely deletes Pod + Volume

Tips to Keep Costs Down

Stop the Pod immediately when done
Save models to Volume — no re-download needed next time
Use Spot instances (50% off). Risk of interruption.
Delete Volume too if you won’t use it for a while (storage is also billed)

Troubleshooting

Model Doesn’t Appear in the List

Check that ComfyUI was restarted
Verify the model file path is correct (/workspace/ComfyUI/models/checkpoints/)
Confirm the download completed

Out of Memory (OOM) Error

Even with RTX 4090’s 24GB, high resolutions or complex workflows can OOM
If using BF16, switch to FP8
Reduce input image size
Use Lightning LoRA to reduce inference steps
If all else fails, try the GGUF version

SSH Connection Fails

Confirm the Pod is running in the RunPod console
Check that your SSH public key is registered in RunPod
Check whether the port is blocked by a firewall

NSFW Support (3D Model Reference Sheets, etc.)

The official model tends to suppress NSFW content. For use cases requiring bare skin, like reference sheets for 3D model base meshes, use a community NSFW variant.

Phr00t/Qwen-Image-Edit-Rapid-AIO

An all-in-one model with VAE/CLIP integrated. Separate NSFW and SFW versions available.

Version	Notes
v18.1-NSFW	Stable, 28.4GB
v19-NSFW	Latest, Lightning 2511 8-step integrated

Download: Phr00t/Qwen-Image-Edit-Rapid-AIO

Setup (Phr00t Version)

Download the Phr00t version instead of the official one.

cd /workspace/ComfyUI/models/checkpoints

# v18.1-NSFW (stable)
huggingface-cli download Phr00t/Qwen-Image-Edit-Rapid-AIO \
  v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors \
  --local-dir ./

Workflow Differences

Since Phr00t’s version has VAE/CLIP integrated, the node setup is simpler.

[Load Checkpoint] Qwen-Rapid-AIO-NSFW-v18.1
       ↓
[Load LoRA] Multiple-Angles-LoRA (strength: 0.8~1.0)
       ↓
[Load Image] Input image
       ↓
[Camera Angle Selector] Select angle
       ↓
[CLIP Text Encode] Prompt
       ↓
[KSampler] steps: 4, cfg: 1.0, sampler: euler_ancestral, scheduler: beta
       ↓
[Save Image]

Recommended Parameters (Phr00t Version)

Parameter	Value
Steps	4~8
CFG	1.0
Sampler	euler_ancestral
Scheduler	beta

Notes

Reports of reduced face consistency in v18+ compared to v16
Worth trying v16 if consistency matters for editing tasks
v18+ has better NSFW quality for text-to-image

The RunPod + ComfyUI environment for Qwen-Image-Edit-2511 can be set up in 1–2 hours
RTX 4090 at $0.34/hour, ~¥2–3 per inference
With Volume reuse, subsequent sessions start in minutes
Smarter to try it in the cloud before spending ¥700,000 on a local setup