How to Install Flux in ComfyUI (Step-by-Step Guide)

Flux.1 is currently the best-quality image generation model available for local use, producing noticeably sharper and more coherent output than SDXL or SD 1.5. Setting it up in ComfyUI is more involved than loading a standard checkpoint — it requires several supporting model files alongside the main model — but the results are worth the effort. This guide covers everything you need from scratch.

Flux.1-dev vs Flux.1-schnell

Black Forest Labs released two variants of Flux.1:

• flux1-dev: The higher-quality variant. Produces the best results but requires more steps (20–30 steps typical) and carries a non-commercial licence. Free to use for personal projects.
• flux1-schnell: A distilled, faster version. Generates good-quality images in 4–8 steps, making it significantly faster. Licensed under Apache 2.0 — commercial use is permitted.

For most personal use cases, start with flux1-schnell — you’ll get results faster and the quality gap at reasonable resolutions is smaller than you might expect. Use flux1-dev when you need the best possible output and are willing to wait longer per image.

Hardware Requirements

Flux is a large model. Be realistic about what your hardware can handle:

• 8GB VRAM: Possible with fp8 or quantised versions (GGUF format). Generation will be slower but functional.
• 12GB VRAM: Comfortable for fp8 Flux. Good balance of quality and speed.
• 16GB+ VRAM: Full bf16 or fp16 precision. Best quality, fastest generation for the GPU class.
• 24GB+ VRAM: Ideal. Runs full-precision Flux without any memory optimisations.

Apple Silicon with 24GB+ unified memory handles Flux well with --force-fp16. 16GB M-series Macs can run quantised versions.

Files You Need

Unlike a standard SD checkpoint (a single file), Flux.1 in ComfyUI’s native format requires three separate components:

1. The Flux model itself (diffusion model / unet)
2. A VAE (Flux has its own VAE — do not use an SD 1.5 or SDXL VAE)
3. Two text encoders: CLIP-L and T5XXL

All files are available from Hugging Face. You’ll need a free account to download them.

Downloading the Files

The Flux Model

Go to huggingface.co/black-forest-labs/FLUX.1-dev (or FLUX.1-schnell for schnell). Download:

• flux1-dev.safetensors (~23GB for bf16) — or for 8GB VRAM systems, look for fp8 versions at huggingface.co/Kijai/flux-fp8

Place this file in: ComfyUI/models/unet/ (create this folder if it doesn’t exist)

The VAE

Go to huggingface.co/black-forest-labs/FLUX.1-dev → Files tab. Download:

• ae.safetensors (~335MB)

Place in: ComfyUI/models/vae/

Text Encoders

You need both CLIP-L and T5XXL. The easiest source for both is huggingface.co/comfyanonymous/flux_text_encoders:

• clip_l.safetensors (~246MB)
• t5xxl_fp16.safetensors (~9.8GB) — or t5xxl_fp8_e4m3fn.safetensors (~5GB) for lower VRAM

Place both in: ComfyUI/models/clip/

Folder Structure Summary

ComfyUI/
  models/
    unet/
      flux1-dev.safetensors        (or flux1-schnell.safetensors)
    vae/
      ae.safetensors
    clip/
      clip_l.safetensors
      t5xxl_fp16.safetensors

Building a Basic Flux Workflow in ComfyUI

Flux requires a different node setup to a standard SD checkpoint — you can’t just swap the model in the default workflow. The easiest approach for beginners is to load a pre-built Flux workflow from the community.

The ComfyUI GitHub repository includes example Flux workflows in the examples/ folder. Alternatively:

1. Open ComfyUI in your browser
2. Click the menu (top right) → Load
3. Load the flux_dev_example.json from the ComfyUI examples folder

The key nodes in a Flux workflow differ from standard SD:

• UnetLoader instead of CheckpointLoader — points to your file in models/unet/
• DualCLIPLoader — loads both CLIP-L and T5XXL simultaneously
• VAELoader — loads ae.safetensors explicitly
• KSampler or SamplerCustomAdvanced — for Flux, use the euler sampler with the simple scheduler

Recommended Settings for Flux

• Steps: 20 for flux1-dev, 4–8 for flux1-schnell
• CFG Scale: 1.0 — Flux is guidance-distilled; higher values produce poor results
• Sampler: euler
• Scheduler: simple (or beta for flux1-dev)
• Resolution: 1024×1024 is the native resolution. Also works well at 832×1216 (portrait) or 1216×832 (landscape)

Using GGUF Quantised Models (Low VRAM)

If you have 8GB VRAM or are on Apple Silicon with limited memory, GGUF quantised Flux models are a good option. They’re smaller and use less VRAM with a modest quality trade-off.

Find these at huggingface.co/city96/FLUX.1-dev-gguf. You’ll also need the ComfyUI-GGUF custom node (installable via ComfyUI Manager) which provides a UnetLoaderGGUF node to load these files.

Place GGUF files in models/unet/ same as the standard models.

Troubleshooting

• CUDA out of memory: Switch to an fp8 or GGUF version of the model, and use the fp8 T5XXL encoder.
• CFG must be 1.0 or you get oversaturated garbage: This is normal for Flux. CFG guidance works differently — always set to 1.0.
• Error loading DualCLIPLoader: Make sure both clip_l.safetensors and t5xxl_fp16.safetensors are in models/clip/ and the node has both set correctly.

Next Steps

With Flux running, explore the wider model landscape — SDXL, Pony, and SD 1.5 all have their uses alongside Flux. See Best Models for ComfyUI in 2026 for a comparison. If you’re new to the node interface generally, the ComfyUI Beginner’s Guide explains the core concepts.

Flux.1 Schnell vs Dev: Which to Use

Flux.1 comes in two main variants. Flux.1 Schnell is the faster version, optimised for 1–4 step generation. It is Apache 2.0 licensed, meaning you can use it freely including commercially. Image quality is excellent for its speed. Flux.1 Dev produces higher quality outputs and handles complex prompts better, but requires more steps (typically 20–30) and is non-commercial licensed. For most personal and testing use cases, Schnell is the better starting point given its speed. Switch to Dev when you need the extra detail for final outputs.

Reducing VRAM Usage with Flux

Flux is significantly more demanding than SDXL. On cards with less than 12 GB VRAM you will need to enable memory optimisations. In your ComfyUI launch command, add --lowvram to reduce peak VRAM usage at the cost of slower generation. For cards with 6–8 GB, use --novae to load the VAE on demand rather than keeping it in VRAM, and consider the GGUF quantised versions of Flux (Q4 or Q8) which cut VRAM requirements substantially. The ComfyUI Manager workflow browser includes pre-built Flux GGUF workflows ready to use.

Testing Your Flux Installation

Once your workflow is set up, queue a generation with a simple prompt such as “a red apple on a wooden table, photorealistic” and click Queue Prompt. Watch the progress bar in the bottom-right of the ComfyUI interface. If the generation completes and an image appears in the preview node, your Flux installation is working correctly. If you see an out-of-memory error, reduce the image dimensions or add the --lowvram flag to your launch command and try again.

For a full index of every ComfyUI guide on Serverman, see the ComfyUI complete guide and hub.