ComfyUI Flux Guide: Setup, Workflows & VRAM Options

What is Flux.1?

Flux.1 is a text-to-image model family created by Black Forest Labs (founded by Robin Rombach, former core member of Stability AI). It delivers exceptional prompt following, text rendering inside images, and natural hand anatomy — areas where Stable Diffusion models traditionally struggle.

Model Versions

Hardware Requirements

Choose Your Setup Path

Flux has multiple packaging formats. Pick the one that matches your hardware:

Path A: FP8 Checkpoint (Recommended for Most Users)

The simplest setup — one file, no extra plugins needed.

Download one model:

Place in ComfyUI/models/checkpoints/.

Workflow: Use a standard Load Checkpoint → CLIP Text Encode → KSampler workflow, just like SD1.5.

Path B: Original Multi-File Setup

Best quality, but requires downloading multiple files:

CLIP models (place in ComfyUI/models/clip/):

VAE (place in ComfyUI/models/vae/):

UNET (place in ComfyUI/models/unet/):

This workflow uses separate loader nodes: UNET Loader, Dual CLIP Loader, and VAE Loader instead of a single Load Checkpoint.

Path C: GGUF (Lowest VRAM)

Requires the ComfyUI-GGUF plugin.

Text-to-Image Workflow

For the FP8 Checkpoint path, the workflow is straightforward:

1. Load Checkpoint → select your flux1-dev-fp8 or flux1-schnell-fp8 model
2. CLIP Text Encode → write your prompt (Flux has excellent text understanding — write naturally)
3. Empty Latent Image → set resolution (recommended: 1024x1024 or similar 1MP resolutions)
4. KSampler → steps: 20 (Dev) or 4 (Schnell), cfg: 1.0, sampler: euler
5. VAE Decode → Save Image

Image-to-Image Workflow

Replace the Empty Latent Image with:

1. Load Image → load your source image
2. VAE Encode → converts the image to latent space

Connect the VAE Encode output to the KSampler's latent input. Adjust denoise (0.3–0.7) to control how much the image changes.

Flux Redux: Image Variation Without Prompts

Flux Redux generates style variants from a reference image — no text prompt required. It needs additional model files:

Flux Ecosystem

Flux has a growing ecosystem of add-ons:

Windows Optimization

If you're running Flux on Windows with limited RAM:

• Increase Windows virtual memory (page file) to at least 32 GB
• Close other applications while generating
• Use the FP8 or GGUF versions to reduce memory pressure

Common Issues and Fixes

Out of memory

• Switch to the FP8 Checkpoint version (8 GB VRAM minimum)
• Or use GGUF Q4 (6 GB VRAM minimum)
• Reduce resolution below 1024x1024

Colors are over-saturated or image looks distorted

• Set cfg to 1.0 — Flux does not use high cfg values like SD1.5

Slow generation speed

• Schnell is 5–10x faster than Dev (4 steps vs 20 steps)
• Use FP8 quantization for faster inference
• See the ComfyUI performance guide

LoRA causing VRAM issues

• Multiple Flux LoRAs can spike VRAM usage significantly
• Consider using the GGUF model when combining LoRAs

Related Guides

• Text to Image — Basic image generation concepts
• LoRA Guide — How to use LoRAs in ComfyUI
• ControlNet Guide — Structural control over generation