Local Image Generation (ComfyUI + FLUX)

A professional-grade image generation pipeline that runs entirely on your own hardware. ComfyUI is a powerful node-based workflow engine for AI image, video, and 3D generation. Paired with FLUX.1 Dev - Black Forest Labs' 12B parameter text-to-image model - you get output that rivals Midjourney and DALL-E, with complete creative control and zero data leaving your machine.

What you get

• Node-based workflow editor - visually construct complex image generation pipelines
• Photorealistic output - FLUX.1 Dev produces stunning 1024x1024 images from text prompts
• 5,000+ community nodes - LoRA, ControlNet, IP-Adapter, AnimateDiff, and more
• Queue-based generation - batch multiple prompts with per-node caching
• Workflow sharing - export/import workflows as JSON from the community
• API mode - integrate with n8n, custom apps, or automation pipelines
• $0/month - after the GPU, every image is free

Architecture

FLUX.1 Dev is a 12B parameter model that needs significant VRAM. Recommended: RTX 4090 24GB for comfortable use, RTX 5090 for faster generation. Can run on 12GB with quantized versions.

Prerequisites

• A GPU with ≥12 GB VRAM (24GB recommended for FLUX at full quality)
• 20 GB free disk for the FLUX model files
• Python 3.10+ or the ComfyUI desktop app
• Git (for custom nodes)

Setup

Option A: Desktop App (Easiest)

1. Download the ComfyUI desktop app from comfy.org
2. Install and launch it
3. Use the built-in model manager to download FLUX.1 Dev

Option B: Manual Install

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt

Download the FLUX model files:

# Create model directories
mkdir -p models/unet models/clip models/vae
 
# Download FLUX.1 Dev (requires HuggingFace login)
# Models go in:
# - models/unet/flux1-dev.safetensors
# - models/clip/clip_l.safetensors
# - models/clip/t5xxl_fp16.safetensors
# - models/vae/ae.safetensors

Launch ComfyUI:

python main.py

Open http://localhost:8188 to access the ComfyUI interface.

Option C: Docker

docker run -d --gpus all -p 8188:8188 \
  --name comfyui \
  -v comfyui_models:/app/models \
  comfyui/comfyui:latest

Using the Basic FLUX Workflow

1. Open http://localhost:8188

2. Load the default workflow or create a new one

3. Add these nodes:

   • Checkpoint Loader → load flux1-dev.safetensors
   • CLIP Text Encoder → enter your prompt (e.g., "a photorealistic cat sitting on a vintage leather chair, warm lighting, depth of field")
   • KSampler → connect model, CLIP, and empty latent
   • VAE Decode → decode the latent to an image
   • Save Image → save the result

4. Click Queue Prompt to generate

Prompt Tips

• FLUX responds well to natural language descriptions
• Add style cues: "photorealistic", "cinematic lighting", "8K", "macro photography"
• Negative prompts work differently in FLUX - use shorter negative prompts than SD models

Advanced Workflows

LoRA Loading

Add a LoRA Loader node between the checkpoint and the model input. LoRA files go in models/loras/.

Image-to-Image

Replace the Empty Latent Image with a VAE Encode node connected to your input image.

ControlNet

Add a ControlNet Loader and Apply ControlNet node for pose/edge guidance.

Video Generation

Install Video Nodes (via ComfyUI Manager) and pair with models like Wan2.1 for AI video.

Cost vs cloud

If you generate 100+ images/month, a 4090 pays for itself in about 2 years versus Midjourney Pro. For power users generating 1000+/month, it pays off in months.

Troubleshooting

• Out of memory → FLUX needs ~20GB VRAM at full precision. Use the FP8 quantized version to fit in 12GB. Lower resolution to 768x768.
• No images showing → Check the console output for errors. The VAE decoder step is often the bottleneck.
• Slow generation → FLUX.1 Dev takes 30-60s per image on a 4090. For faster results, use FLUX.1 Schnell (4-step distilled version).
• Missing model files → Use the ComfyUI Manager node to download models from within the interface.
• CORS errors in API mode → Set --listen 0.0.0.0 and configure the API settings in extra_model_paths.yaml.

Swap components

• Faster generation → Use FLUX.1 Schnell (4-step distilled, 4x faster)
• Video generation → Add Wan2.1 or HunyuanVideo nodes
• Alternative UI → Try SwarmUI for a simpler interface on top of ComfyUI
• Lower VRAM → Use SDXL or SD3.5 for 8GB cards
• n8n integration → Use ComfyUI's API mode with n8n for automated generation pipelines