DarokarAmey/Image-Manipulation-Comfyui

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

🧠 ComfyCPU Studio (2vCPU Edition)

A lightweight ComfyUI CPU setup running on:

2 vCPUs + 16GB RAM No GPU. Just optimization.

I built this because I genuinely can’t afford GPU upgrades right now. So instead of waiting for better hardware, I optimized everything to run properly on CPU.

This Space focuses on practical workflows that actually work on limited hardware.

You are free to make new workflows and test this space's limitation.

If you want to see whats happening in the workflow, click on "Console" in bottom left.

PS - If it breaks/crashes you just have to click on - "Manager" top right, then on - "Restart" in red box. It should then reboot the Comfyui properly.

🚀 Core Workflows

All workflows below are embedded inside the PNG images.

👉 Just drag the image into ComfyUI and the full workflow loads automatically.

Do NOT re-save the PNG or compress it — metadata will be lost.

🖼 1️⃣ SD 1.5 Image Generation (CPU Optimized)

512×512 Stable Diffusion 1.5 generation tuned for 2vCPU.

• —15–18 steps
• —Euler sampler
• —CFG 6–7
• —Batch size = 1
• —No high-res fix

⏳ Expect ~1–3 minutes per image (this is CPU life).

Drag this into ComfyUI:

[image]

🔍 2️⃣ Image Upscale (RealESRGAN + ClearReality)

Enhance low-resolution images using lightweight upscalers.

Perfect for:

• —AI images
• —Old photos
• —Small thumbnails
• —Light detail recovery

Uses:

• —RealESRGAN_x4plus
• —4x-ClearRealityV1

Drag this into ComfyUI:

[image]

✂️ 3️⃣ Background Remover (Rembg / U2Net)

Clean background removal using ONNX U2Net.

Great for:

• —Product photos
• —Profile pictures
• —PNG cutouts
• —Editing pipelines

Fast and very stable on CPU.

Drag this into ComfyUI:

[image]

💡 Why CPU Only?

Because I’m broke.

Simple.

No GPU. No paid hardware upgrades. Just patience and optimization.

Instead of chasing expensive infrastructure, I focused on:

• —Stable workflows
• —Clean outputs
• —Real-world use cases
• —CPU-safe configuration

This project proves you can still build useful AI tools without a GPU.

⚙️ Hardware & Limits

Running on:

• —2 vCPU
• —16GB RAM
• —Docker Space

✅ What works well

• —512×512 diffusion
• —Face restoration
• —Background removal
• —Upscaling
• —Light batch processing

❌ What does NOT work well

• —SDXL
• —1024×1024 generation
• —Multiple ControlNets
• —High-res fix loops
• —Batch size > 1

Everything here is tuned for CPU survival.

🛠 Recommended Settings

For best performance:

• —Resolution: 512×512
• —Steps: 15
• —CFG: 6–7
• —Sampler: Euler
• —Batch size: 1
• —Disable high-res fix

If it feels slow:

• —Reduce steps to 12
• —Try 448×448

🧩 Experimental / In Progress

I’m slowly experimenting with:

• —Face restoration (GFPGAN / CodeFormer)
• —Portrait enhancer pipeline
• —Short video frame upscaling
• —Cleaner UI presets

As long as this stays on 2vCPU, everything will remain CPU-optimized.

🧠 Future Plans

If (very big if) I eventually get access to a GPU:

• —SDXL support
• —Better ControlNet workflows
• —IPAdapter stylization
• —Automation pipelines
• —Cleaner public UI version

For now, this is a CPU lab.