Image ToolsIntermediate

How to Run Flux.1 Dev in ComfyUI

Download the Flux.1 Dev model, its text encoders and VAE, place each file correctly, and load the default Flux workflow to generate your first Flux image.

10 minIntermediate

Flux.1 Dev is a strong open-weight text-to-image model from Black Forest Labs. It is split into several files that go into different folders, which trips up most newcomers. Once the pieces are placed correctly, ComfyUI's built-in Flux template just works. This guide walks through every file and where it lives.

What you need

ComfyUI updated to a recent version (Flux support is built in)
A GPU with 12 GB VRAM for the full model, or use the fp8 build on 8 GB
A Hugging Face account to accept the Flux Dev license and download
Around 30 GB of free space for all the files combined

Step 1: Download the four pieces

Flux needs the diffusion model itself, two text encoders, and a VAE. Grab these files and note which folder each one belongs in. The fp8 single-file version of the encoder is the easy path for limited VRAM.

File	Goes in folder
`flux1-dev.safetensors`	`models/unet`
`clip_l.safetensors`	`models/clip`
`t5xxl_fp16.safetensors`	`models/clip`
`ae.safetensors`	`models/vae`

Tight on VRAM?

Swap t5xxl_fp16 for t5xxl_fp8_e4m3fn, and use the flux1-dev-fp8 checkpoint. The fp8 versions roughly halve memory use with a small quality cost, which keeps Flux usable on an 8 GB card.

Step 2: Place each file in the right folder

This is where most failures happen. The diffusion model is a unet, not a checkpoint, so it does not go in models/checkpoints. Double-check the layout against the tree below.

ComfyUI/models

models/

unet/

flux1-dev.safetensors

clip/

clip_l.safetensors

t5xxl_fp16.safetensors

vae/

ae.safetensors

Four files, three folders. The unet folder is the one people forget.

Step 3: Load the Flux template

In ComfyUI open the workflow menu and choose a Flux template (in recent builds, Workflow > Browse Templates > Flux). This drops in a ready graph wired for the unet loader, dual CLIP loader, and VAE.

Step 4: Select your files in the loaders

Click each loader node and pick the matching file: flux1-dev in the UNET loader, clip_l plus t5xxl in the DualCLIPLoader (type set to flux), and ae.safetensors in the VAE loader. If a dropdown is empty, the file is in the wrong folder.

ComfyUI - Flux loader nodes

Load Diffusion Model DualCLIPLoader

unet: flux1-dev.safe.. clip1: t5xxl_fp16

weight_dtype: default clip2: clip_l

type: flux

Load VAE EmptyLatentImage

vae: ae.safetensors 1024 x 1024

Step 5: Set Flux sampling values and generate

Flux Dev likes around 20 to 25 steps with a guidance value near 3.5 and the euler sampler. Type a prompt, set the latent to 1024 by 1024, and click Queue Prompt. The first run loads the large T5 encoder, so give it a minute.

Result: a Flux Dev render at native 1024 resolution. Flux follows long, descriptive prompts well and renders legible text, so write full sentences rather than keyword soup.