🔥 QWEN IMG2IMG ULTRA GRAND FINAL

✨ Overview

This workflow is a high-quality Img2Img pipeline built around Qwen Image, designed to push:

• 🎯 image fidelity

• 🧠 prompt understanding

• 🧩 structural consistency

• 🎨 final visual quality

🎯 Workflow goals

✔ maximum image control
✔ excellent prompt interpretation
✔ premium visual output
✔ powerful and clean upscaling
✔ modular & easy customization

🎨 Use cases

This workflow is perfectly suited for:

• 🧍 realistic images

• 🎭 stylized art

• 🧑‍🎨 character refinement

• ✨ image enhancement

• 🎬 cinematic renders

• 🔍 controlled img2img generations

⚙️ What this workflow does

This pipeline includes:

✅ Qwen Image as the core generative model
✅ Full Img2Img workflow with strong structure preservation
✅ Smart denoise control for the right balance between creativity and fidelity
✅ Ultra-sharp final upscale using SeedVR2
✅ Optional face detection & segmentation for cleaner results
✅ Clean node logic — easy to read and modify
✅ Optimized for stability and repeatability

👉 This workflow is designed to produce consistent results, not random outputs.

🏆 Why this workflow stands out

✔ exceptional prompt comprehension
✔ better composition consistency than classic SD pipelines
✔ strong identity preservation
✔ cleaner textures
✔ improved line stability
✔ fewer upscale artifacts
✔ more predictable results

🎯 Ideal for creators who want quality + control, not chaos.

📦 Required Models (Direct Downloads)

🧠 Qwen Image

UNet GGUF
➡️ Recommended: Q8_0 (best quality / speed balance)
👉 https://huggingface.co/unsloth/Qwen-Image-2512-GGUF/resolve/main/qwen-image-2512-Q8_0.gguf?download=true

🔹 Lightweight alternative (lower VRAM):
👉 https://huggingface.co/unsloth/Qwen-Image-2512-GGUF/resolve/main/qwen-image-2512-Q4_K_M.gguf?download=true

VAE
👉 https://huggingface.co/Qwen/Qwen-Image-2512/resolve/main/vae/qwen_image_vae.safetensors?download=true

(Fallback if needed)
👉 https://huggingface.co/Qwen/Qwen-Image/resolve/main/vae/qwen_image_vae.safetensors?download=true

Text Encoder / CLIP
👉 https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/resolve/main/split_files/text_encoders/qwen_2.5_vl_7b_fp8_scaled.safetensors?download=true

🔹 FP16 version available in the same repository if required.

🚀 SeedVR2 Upscaler

DiT GGUF (sharp version recommended)
👉 https://huggingface.co/cmeka/SeedVR2-GGUF/resolve/main/seedvr2_ema_7b_sharp-Q4_K_M.gguf?download=true

🔹 Other quantizations available here:
👉 https://huggingface.co/AInVFX/SeedVR2_comfyUI/tree/main/gguf

VAE
👉 https://huggingface.co/AInVFX/SeedVR2_comfyUI/resolve/main/vae/ema_vae_fp16.safetensors?download=true

👤 Face & Mask Models

YOLOv8 Face Detector
👉 https://huggingface.co/Bingsu/adetailer/resolve/main/face_yolov8m.pt?download=true

SAM — Segment Anything Model
👉 https://dl.fbaipublicfiles.com/segment_anything/sam_vit_b_01ec64.pth

📁 Models Installation Path

models/
├─ unet/
│  └─ Qwen/
│     └─ qwen-image-2512-Q8_0.gguf
├─ vae/
│  ├─ qwen_image_vae.safetensors
│  └─ ema_vae_fp16.safetensors
├─ clip/
│  └─ qwen_2.5_vl_7b_fp8_scaled.safetensors
├─ gguf/
│  └─ seedvr2_ema_7b_sharp-Q4_K_M.gguf
├─ sam/
│  └─ sam_vit_b_01ec64.pth
└─ ultralytics/
   └─ face_yolov8m.pt

💪CREATED AND BATTLE TESTED WITH 🚀

This workflow runs rock-solid and ultra-fast on my beast setup:

- CPU: Intel Core i9-14900KF

- GPU: ASUS RTX 5090 ASTRAL

- RAM: 128GB DDR5 8000MHz Corsair Vengeance

- Motherboard: ASUS ROG Maximus Z790 Extreme

- PSU: ASUS ROG STRIX 1275W Platinum

- Case: Hyte Y70 Touch

- Cooling: Asus Ryujin III Extreme 360 ARGB (Noctua edition) AIO Watercooling

- Storage: 30TB NVMe M.2 (high-speed array)

Zero crashes, buttery smooth generations & upscales even at high res / batch sizes.

😎 Casual Tested With My 2nd Config 😎

Yeah, it runs smooth on my older beast, no need for a 5090 to enjoy this workflow!

- CPU: Intel Core i9-9900K

- GPU: ASUS TUF NVIDIA RTX 3080 Ti

- RAM: 128GB DDR4 5200MHz

- Motherboard: ASUS TUF Z390-PLUS GAMING (WiFi)

- Cooling: NZXT Kraken X63

- PSU: Gigabyte 750W 80+ Gold

- Storage: 8TB NVMe M.2

Expect lower speed generations and executions.

⚠️ Important Notes

• 🔧 Designed for ComfyUI

• 💾 Recommended GPU: 12–16 GB VRAM or more

• 🖼️ Final results strongly depend on input image quality

• 🔄 Always restart ComfyUI after installing new models

• 🆓 Workflow provided as-is, free to use and modify

⚡Personal BackEnd Setup For the 1st and 2nd Config 🚀😏

My Backend Setup

- Python 3.10.6

- PyTorch 2.7.0 + CUDA 12.8

- Flash Attention 2 (custom build)

- SageAttention 2.1.1

- xFormers (custom build)

- DeepSpeed 0.16.5

- Triton Windows 3.3.0.post19

- torchao + torchsde

- huggingface_hub with hf_transfer (for super-fast downloads)

❤️ Message from the Creator

I spent many hours building, testing, refining, and optimizing this workflow with one clear goal:

👉 Make it powerful, stable, and accessible to everyone.

If this workflow helps you, inspires you, or saves you time, I would truly appreciate it if you could:

• ❤️ leave a like / buzz

• ⭐ follow my Civitai page

• 🎨 share your creations made with this workflow

Seeing what you create with it is honestly the best reward.

👉 https://civarchive.com/user/EKKIVOK

Thank you for your support 🙏
— EKKIVOK

⚠️ Important Disclaimer & Responsible Use

This workflow is provided for creative, artistic, and personal experimentation purposes only.

- Do not create or generate images of real people without their explicit consent.

- Do not use it for deepfakes, non-consensual explicit content, impersonation, harassment, deception, or any harmful/illegal purpose.

- Respect copyright, trademarks, privacy laws, and all applicable regulations in your country.

- When sharing generated images publicly (including on Civitai), clearly indicate they are AI-generated and follow Civitai's [Content Rules & TOS](https://civarchive.com/content/tos).

By using this workflow, you agree to use it in a lawful, ethical, and responsible manner.

Thank you for keeping AI creation positive and respectful ✨