Translated from the original Japanese article
Tech 7 min read

Running the NSFW Version of Qwen-Image-Edit on RunPod

RunPod Qwen ComfyUI Image Generation Experiment

I wanted to generate 3-view reference sheets for a 3D model base mesh, so I tried running the NSFW variant of Qwen-Image-Edit (Phr00t AIO) on RunPod. This comes after checking local specs and putting together a setup guide — this time was the first real run. Things went sideways: RTX 4090 (24GB) didn’t work, and I finally got it running on RTX 5090 (32GB).

What I Tried Today

Failure on RTX 4090 (24GB VRAM)

Environment

  • RunPod RTX 4090 (24GB VRAM)
  • Template: runpod/comfyui:latest
  • ComfyUI path: /workspace/runpod-slim/ComfyUI/

Why It Failed

  • Phr00t AIO NSFW v18.1 (28GB) doesn’t fit in 24GB VRAM
  • --lowvram option didn’t help
  • FP8 version (20GB) requires VAE/Text Encoder separately, making the setup complex

Models I Tried

1. Phr00t AIO NSFW v18.1 (28GB)

  • Path: models/checkpoints/v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors
  • Result: Failed — 28GB doesn’t fit in 24GB VRAM, froze during load
  • --lowvram didn’t help

2. 1038lab FP8 (20GB)

  • Path: models/checkpoints/fp8/Qwen-Image-Edit-2511-FP8_e4m3fn.safetensors
  • Result: VAE error — this file is diffusion model only; VAE/Text Encoder required separately
  • Error: ERROR: VAE is invalid: None

What I Learned

1. Model Configuration Differences

Phr00t AIO Version (All-In-One)

  • VAE/CLIP integrated
  • Self-contained in one file
  • But 28GB doesn’t fit in 24GB VRAM

Official FP8 / 1038lab FP8

  • Diffusion model only
  • Requires separately:
    • VAE: qwen_image_vae.safetensors
    • Text Encoder: qwen_2.5_vl_7b_fp8_scaled.safetensors

2. Required Custom Node

Standard CLIPTextEncode won’t work — you need TextEncodeQwenImageEditPlus.

This is Phr00t’s modified version of ComfyUI’s standard nodes_qwen.py. Install it with:

# Run from the ComfyUI directory
cd comfy_extras
wget https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/resolve/main/fixed-textencode-node/nodes_qwen.py -O nodes_qwen.py

Source: https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/tree/main/fixed-textencode-node

3. Workflow Differences

Standard SD/FLUX workflow (CheckpointLoader → LoRA → KSampler) doesn’t work. A Qwen-specific node setup is required.

VRAMRecommended SetupNotes
12GBGGUF Q2_K (7.4GB)Lightest, quality somewhat reduced
16GBGGUF Q4_K_M (13.3GB)Balanced, recommended
24GBGGUF Q5_K_M (15.1GB) or Official FP8 split setupHigh quality
32GBPhr00t AIO NSFW (28GB)Full size, best quality

Options for RTX 4090 (24GB VRAM)

Q4_K_M is about 13GB, fits comfortably in 24GB VRAM. NSFW-capable GGUF quantization.

Required custom node:

cd /workspace/ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
pip install --upgrade gguf

Required files:

models/unet/
  └── Qwen-Rapid-AIO-NSFW-v18.1-Q4_K_M.gguf

models/text_encoders/
  └── Qwen2.5-VL-7B-Instruct-abliterated-Q4_K_M.gguf
  └── mmproj-xxx.gguf (must be in the same directory)

models/vae/
  └── pig_qwen_image_vae_fp32-f16.gguf

Download:

pip install huggingface_hub

# GGUF main model
cd /workspace/ComfyUI/models/unet
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Phil2Sat/Qwen-Image-Edit-Rapid-AIO-GGUF', 'Qwen-Rapid-AIO-NSFW-v18.1-Q4_K_M.gguf', local_dir='./')"

# Text Encoder (abliterated version)
cd /workspace/ComfyUI/models/text_encoders
python3 -c "from huggingface_hub import snapshot_download; snapshot_download('Phil2Sat/Qwen-Image-Edit-Rapid-AIO-GGUF', allow_patterns='Qwen2.5-VL-7B-Instruct-abliterated/*', local_dir='./')"

# VAE
cd /workspace/ComfyUI/models/vae
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('calcuis/pig-vae', 'pig_qwen_image_vae_fp32-f16.gguf', local_dir='./')"

Download source: https://huggingface.co/Phil2Sat/Qwen-Image-Edit-Rapid-AIO-GGUF

Option B: Official FP8 Split Setup

SFW version. Loads VAE, Text Encoder, and Diffusion Model separately.

Required files:

models/diffusion_models/
  └── qwen_image_edit_fp8_e4m3fn.safetensors

models/vae/
  └── qwen_image_vae.safetensors

models/text_encoders/
  └── qwen_2.5_vl_7b_fp8_scaled.safetensors

models/loras/
  └── Qwen-Image-Lightning-4steps-V1.0.safetensors (optional, for 4-step generation)

Download:

pip install huggingface_hub

cd /workspace/ComfyUI/models

# Diffusion Model
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Comfy-Org/Qwen-Image-Edit_ComfyUI', 'qwen_image_edit_fp8_e4m3fn.safetensors', local_dir='./diffusion_models/')"

# VAE
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Comfy-Org/Qwen-Image_ComfyUI', 'qwen_image_vae.safetensors', local_dir='./vae/')"

# Text Encoder
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Comfy-Org/Qwen-Image_ComfyUI', 'qwen_2.5_vl_7b_fp8_scaled.safetensors', local_dir='./text_encoders/')"

# Lightning LoRA (optional)
python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Comfy-Org/Qwen-Image-Edit_ComfyUI', 'Qwen-Image-Lightning-4steps-V1.0.safetensors', local_dir='./loras/')"

Download sources:

Workflow reference: https://docs.comfy.org/tutorials/image/qwen/qwen-image-edit

Workflows

Models

Custom Nodes

RunPod Operation Notes

Billing

  • RTX 4090: $0.59–0.6/hr (as of January 2026)
  • RTX 5090: $0.9/hr (as of January 2026)
  • Volume retention (when stopped): $0.022/hr ($0.53/day, ~$16/month)

How to Stop a Pod

  • Stop: GPU billing pauses, Volume kept, can resume quickly
  • Delete + Keep Volume: Pod deleted, Volume kept, need to create new Pod next time
  • Terminate: Completely deleted

This Session’s Volume Contents

/workspace/runpod-slim/ComfyUI/models/
├── checkpoints/
│   ├── v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors (28GB)
│   └── fp8/Qwen-Image-Edit-2511-FP8_e4m3fn.safetensors (20GB)
└── loras/
    └── Multiple-Angles-LoRA/

Successful Run (Confirmed 2026/01/24)

Input image:

Input image

ComfyUI generation screen:

ComfyUI generation screen

Environment

  • GPU: RTX 5090 (32GB VRAM)
  • Template: ashleykza/comfyui:cu128-py312-v0.10.0
  • PyTorch: 2.9.1+cu128
  • ComfyUI: 0.10.0

Model

  • Phr00t AIO NSFW v18.1 (28GB)
  • Path: models/checkpoints/v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors

Workflow

Settings

  • steps: 4
  • cfg: 1.0
  • sampler: sa_solver
  • scheduler: beta
  • Output size: 768x768 (for testing; 1024 recommended for production)

Default prompt output:

Default prompt output

Prompt (for 3D Model Reference Sheet)

Positive:

front view, full body, nude, same character, simple flat colors, white background

Negative:

blurry, bad anatomy, deformed, low quality, text, watermark, clothes

Reference sheet prompt output (no mosaic in the actual output):

Reference sheet prompt output

Generating a 3-View Sheet

  • Generate separately for front view → side view → back view
  • Use the same input image for all three

Generation Speed

  • Blazing fast (4 steps in a few seconds)

Steps for RTX 5090 (32GB VRAM)

1. Create Pod

  • GPU: RTX 5090 (32GB VRAM)
  • Template: ashleykza/comfyui:cu128-py312-v0.10.0 (RTX 5090 compatible)
  • Disk: default is fine

Note: RTX 5090 requires PyTorch 2.8+. Standard ComfyUI templates don’t support it.

2. Download Model (Web Terminal)

Open Web Terminal and run:

pip install huggingface_hub

cd /workspace/ComfyUI/models/checkpoints

python3 -c "from huggingface_hub import hf_hub_download; hf_hub_download('Phr00t/Qwen-Image-Edit-Rapid-AIO', 'v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors', local_dir='./')"

Note: 28GB, takes about 5 minutes.

3. Load Workflow

  1. Open in browser: https://huggingface.co/Phr00t/Qwen-Image-Edit-Rapid-AIO/blob/main/Qwen-Rapid-AIO.json
  2. Click “Download” to save locally
  3. Drag and drop the JSON file into ComfyUI (Port 3000)

4. Run

  • Checkpoint: select v18/Qwen-Rapid-AIO-NSFW-v18.1.safetensors
  • Upload an image
  • Set prompt and run