How to Use Flux in ComfyUI (2026): Which Model, Files, Settings, LoRAs and Fixes

Flux is the image model family from Black Forest Labs, and it is still one of the best ways to make sharp, prompt-accurate images on your own PC. The catch is that “Flux” now means several different models with different licences, file layouts and VRAM needs, and old tutorials mix them up.

This guide shows how to use Flux the easy way in 2026. You will learn which Flux model to pick, exactly which files go in which ComfyUI folder, how to run your first image, the settings that matter, how to add LoRAs and what to do when something breaks.

What Flux Is

Flux is a family of text-to-image models made by Black Forest Labs. You type a prompt, and the model turns it into an image. Newer versions can also edit images and use reference pictures.

Flux became popular because it follows long prompts closely, handles hands and faces better than older Stable Diffusion models and can write readable text inside images.

Unlike older checkpoints, most Flux models are split into separate parts: a diffusion model, one or more text encoders and a VAE. That is why the setup looks different from SD 1.5 or SDXL, and why “missing file” errors are so common.

You can run Flux locally in ComfyUI, which is free and private, or through paid cloud APIs. This guide covers both, with the main focus on ComfyUI. If you have never used ComfyUI, start with our ComfyUI install guide and the ComfyUI beginners guide.

Which Flux Model Should You Use?

Here is the short version of the Flux line-up as of November 2026.

Model	Best for	Licence	Notes
FLUX.1 schnell	Fast images, commercial use	Apache 2.0	12B parameters, about 4 steps
FLUX.1 dev	High quality, huge LoRA library	Non-commercial	12B parameters, 20 or more steps
FLUX.1 pro	Top FLUX.1 quality	API only	No downloadable weights
FLUX.2 Dev	Up to 4MP images, editing, up to 10 reference images, text in images	Check model card	Big model, has an FP8 template for less VRAM
FLUX.2 Klein 4B	Fast local use on mid-range GPUs	Check model card	Base and 4-step distilled versions
FLUX.3 Klein	Newest open Flux, reference images, low VRAM options	FLUX Kommunity License	Needs ComfyUI 0.41.0 or newer
FLUX 3 Image / Max / Video	No local GPU	Paid API	ComfyUI partner nodes

Simple rule of thumb:

• You want to sell or publish images commercially and run locally: FLUX.1 schnell, or check the FLUX.3 Klein licence terms for your use.
• You want the most LoRAs and tutorials: FLUX.1 dev. Most Flux LoRAs on Civitai and Hugging Face are made for it.
• You have 8 to 12 GB of VRAM: FLUX.2 Klein 4B distilled or FLUX.3 Klein (int8 or nvfp4 files).
• You want editing with several reference images: FLUX.2 Dev or FLUX.3 Klein.

Always read the licence on the Hugging Face model page before using images in paid work. Licences differ between models, and the “dev” licence does not allow commercial use of the model itself.

What You Need Before You Start

• ComfyUI, updated: the desktop app or portable build. FLUX.3 Klein needs version 0.41.0 or newer.
• An NVIDIA GPU is best: FLUX.2 Klein 4B distilled used about 8.4 GB of VRAM in Comfy’s own test, so 8 to 12 GB cards are fine for the small models. FLUX.1 dev at full precision wants 24 GB, but FP8 and GGUF files bring that down.
• System RAM: 32 GB is comfortable. 16 GB works with smaller files.
• Disk space: plan for 20 to 40 GB for one Flux model with its text encoders.
• A Hugging Face account: FLUX.1 dev is a gated model. You must log in and accept the licence on its page before the download works.

Method 1: Use a Flux Template (Easiest)

ComfyUI ships ready-made Flux workflows. This is the fastest way to get your first image, and it avoids wiring mistakes.

1. Update ComfyUI. In the desktop app, use the update option in the menu. In the portable build, run the update script in the update folder.
2. Open the template browser. Click Workflow > Browse Templates and search for “flux”.
3. Pick a template. Good starting points are the FLUX.1 dev or schnell text-to-image templates, image_flux2_text_to_image, image_flux2_fp8 for lower VRAM, or the FLUX.3 Klein template.
4. Download missing models. When the workflow opens, ComfyUI shows a list of missing model files with download buttons. Download each one. The template already knows the right folder.
5. Edit the prompt. Find the text encode node, replace the example prompt with your own, and leave the other settings as they are for now.
6. Click Run. The first run is slow because models load into memory. Later runs are much faster.

If you want to understand how templates and shared workflow files work, our ComfyUI workflows guide explains loading, saving and fixing them.

Method 2: Set Up FLUX.1 Dev or Schnell by Hand

Doing it by hand helps you understand Flux, and it is useful when a template download fails.

Step 1: Download the files

File	Folder	Purpose
flux1-dev.safetensors or flux1-schnell.safetensors	ComfyUI/models/diffusion_models/	The main image model
clip_l.safetensors	ComfyUI/models/text_encoders/	First text encoder
t5xxl_fp16.safetensors (or t5xxl_fp8_e4m3fn.safetensors for low VRAM/RAM)	ComfyUI/models/text_encoders/	Second text encoder (T5)
ae.safetensors	ComfyUI/models/vae/	Turns the result into pixels

The model files are on the Black Forest Labs Hugging Face pages. The text encoders are in the comfyanonymous/flux_text_encoders repository. Older guides put text encoders in models/clip; that still works, but text_encoders is the current folder.

There are also single-file FP8 checkpoints of FLUX.1 dev and schnell that bundle everything. Those go in models/checkpoints/ and load with a normal Load Checkpoint node, which is simpler if you are short on VRAM.

Step 2: Build the workflow

1. Load Diffusion Model: select flux1-dev.safetensors. Leave weight_dtype on default, or pick an fp8 option to save VRAM.
2. DualCLIPLoader: set clip_name1 to t5xxl_fp16 (or fp8), clip_name2 to clip_l and type to flux.
3. Load VAE: select ae.safetensors.
4. CLIP Text Encode: connect it to the DualCLIPLoader and write your prompt.
5. FluxGuidance (dev only): put it after the text encode node. A guidance of 3.5 is the usual starting point.
6. Empty Latent Image (or EmptySD3LatentImage): set 1024 x 1024 to start.
7. KSampler: cfg 1.0, sampler euler, scheduler simple. Use 20 to 30 steps for dev and 4 for schnell.
8. VAE Decode and Save Image at the end.

You do not need a negative prompt with Flux. If your workflow has a negative prompt box, leave it empty. With cfg at 1.0 it is ignored anyway.

Step 3: Run and check

Click Run. If the image looks grey or noisy, check that the VAE is ae.safetensors and that the DualCLIPLoader type is set to flux. Those two mistakes cause most broken first images.

Method 3: FLUX.2 Klein and FLUX.3 Klein for Low VRAM

The Klein models are the smaller, faster members of the newer Flux generations. They are the best choice for most home PCs.

FLUX.2 Klein 4B

FLUX.2 Klein 4B comes in two versions:

• Distilled: 4 steps. In Comfy’s test on an RTX 5090 it made an image in about 1.2 seconds using about 8.4 GB of VRAM.
• Base: more steps and slower (about 17 seconds in the same test, 9.2 GB VRAM), but better for fine-tuning and LoRA training.

Load it from the FLUX.2 Klein template and let ComfyUI fetch the files. Use the distilled model for everyday images.

FLUX.3 Klein

FLUX.3 Klein is the newest open Flux model. Key facts from the ComfyUI docs:

• Open weights under the FLUX Kommunity License.
• Needs ComfyUI 0.41.0 or newer.
• Two versions: Full (around 50 steps, cfg 3.5) and Distilled (4 to 16 steps, cfg 1).
• Uses a Qwen3-VL-4B text encoder and the FLUX.2 VAE.
• Has a built-in Qwen3.5 9B prompt upsampler that rewrites short prompts into detailed ones. It is on by default and can be turned off.
• Comes in bf16, int8 and nvfp4 versions. Use int8 or nvfp4 on smaller cards; nvfp4 is aimed at RTX 50-series GPUs.
• Supports reference images, which you mention in the prompt as <ref_image_0>, <ref_image_1> and so on.

For exact file names, folders and an 8 GB VRAM recipe, see our full FLUX 3 Image Klein in ComfyUI guide.

GGUF files for very low VRAM

If even the small models run out of memory, use GGUF quantized versions of Flux with the ComfyUI-GGUF custom node. Our GGUF in ComfyUI guide covers installing the node, which quant to pick and where the files go.

Method 4: Use Flux Without a GPU

If your PC has no suitable GPU, you can still use Flux:

• ComfyUI partner nodes: ComfyUI includes FLUX 3 Image, FLUX 3 Max and FLUX 3 Video nodes. They run on Black Forest Labs servers, you sign in to a Comfy account and pay with credits per generation. Your PC only sends the prompt and receives the image.
• Black Forest Labs API: developers can call the official API directly and pay per image.
• Rented cloud GPUs: run ComfyUI on a rented GPU by the hour and follow the local methods above.

Check current prices on the provider’s pricing page before big batches, because API prices change often.

Best Flux Settings

Model	Steps	CFG	Guidance	Sampler / scheduler
FLUX.1 dev	20 to 30	1.0	FluxGuidance 3.5 (2 to 4 range)	euler / simple
FLUX.1 schnell	4	1.0	Not used	euler / simple
FLUX.2 Klein 4B distilled	4	1.0	Template default	Template default
FLUX.3 Klein Full	about 50	3.5	Template default	Template default
FLUX.3 Klein Distilled	4 to 16	1.0	Template default	Template default

A few more tips:

• Resolution: 1024 x 1024 is a safe start for FLUX.1. Wide formats like 1344 x 768 work well too. FLUX.2 Dev can go up to about 4 megapixels.
• Do not raise cfg on distilled models. cfg above 1 on schnell or distilled Klein usually burns colours and adds artefacts.
• Lower FluxGuidance for realism. Around 2 to 2.5 often looks more natural for photos; 3.5 to 4 follows the prompt more strictly.
• Keep the seed fixed while you tweak prompts, so you can see what each change does.

How to Write Flux Prompts

Flux understands normal sentences much better than older models. Write prompts like you are describing a photo to a friend.

• Start with the subject: “A smiling woman in a yellow raincoat holding a coffee cup”.
• Add the setting: “standing on a rainy city street at night, neon shop signs behind her”.
• Add the style and camera: “cinematic photo, shallow depth of field, 35mm lens”.
• Put text in quotes: “a cafe sign that says “OPEN 24/7″”. Flux renders short quoted text well.
• Skip keyword soup: long lists like “masterpiece, best quality, 8k” add little with Flux.

Example prompt:

A cheerful young woman in a lime green hoodie sitting at a wooden desk with a laptop,
warm morning sunlight through a window, small plants on the shelf behind her,
a mug with the text "FLUX" on the desk, realistic photo, soft focus background

With FLUX.3 Klein, short prompts also work because the built-in upsampler expands them. Turn the upsampler off when you want the model to follow your exact words.

Using LoRAs With Flux

LoRAs let you add a style, character or concept to Flux without changing the base model.

1. Download a LoRA made for your exact Flux version, for example FLUX.1 dev. A FLUX.1 LoRA will not work on FLUX.2 or FLUX.3.
2. Put it in ComfyUI/models/loras/ and press R in ComfyUI to refresh the lists.
3. Add a LoraLoaderModelOnly node between Load Diffusion Model and the KSampler.
4. Set strength to 0.8 to 1.0 and add the LoRA’s trigger word to your prompt.

Our step-by-step guide to using LoRAs in ComfyUI covers stacking several LoRAs and strength settings. To make your own, see how to train a LoRA with Kohya_ss. For editing existing images instead of generating new ones, read the FLUX.1 Kontext Dev guide.

Troubleshooting

Problem	Likely cause	Fix
Model not in the dropdown	Wrong folder or list not refreshed	Check the folder table above, then press R or restart ComfyUI
Download fails with 401 or 403	Gated FLUX.1 dev model	Log in to Hugging Face and accept the licence on the model page first
Grey, noisy or washed-out image	Wrong VAE or wrong CLIP type	Use ae.safetensors and set DualCLIPLoader type to flux
Out of memory (CUDA OOM)	Model too big for your GPU	Use FP8 files, t5xxl_fp8_e4m3fn, a Klein model or GGUF; lower the resolution
Very slow, PC freezes	Not enough system RAM, heavy swapping	Use fp8 text encoders, close other apps, add RAM or use smaller models
Burnt, over-saturated colours	cfg too high on a distilled model	Set cfg to 1.0 for schnell and distilled Klein
Template shows red “missing node” boxes	ComfyUI too old	Update ComfyUI; FLUX.3 Klein needs 0.41.0+
LoRA has no effect	LoRA made for another Flux version	Match the LoRA to your base model and use LoraLoaderModelOnly
Text in images is garbled	Text too long or not quoted	Keep text short and in quotes; try FLUX.2 Dev or FLUX.3 Klein

My Take

If you are new, do not start by hand-wiring nodes. Update ComfyUI, open a Flux template and let it download the files. You will have a working image in minutes, and you can study the wiring afterwards.

For most home PCs in late 2026, FLUX.3 Klein Distilled or FLUX.2 Klein 4B distilled is the sweet spot of speed and quality. Keep FLUX.1 dev installed if you want access to the enormous LoRA library, and use schnell when you need a permissive licence.

And if your laptop has no GPU at all, the FLUX 3 partner nodes give you the latest Flux quality inside the same ComfyUI interface for a per-image fee.

FAQ

Is Flux free to use?

The open-weight models are free to download and run locally. FLUX.1 schnell uses the Apache 2.0 licence. FLUX.1 dev is free for non-commercial use. FLUX.3 Klein uses the FLUX Kommunity License. The pro, Max and API models are paid.

Can I use Flux images commercially?

It depends on the model. Schnell’s Apache 2.0 licence allows it. For dev and Klein models, read the licence on the model page, because terms differ and can change.

How much VRAM do I need for Flux?

About 8 GB is enough for FLUX.2 Klein 4B distilled and the int8 or nvfp4 FLUX.3 Klein files. FLUX.1 dev at full precision wants around 24 GB, but FP8 and GGUF versions run on 12 GB and below.

Does Flux need a negative prompt?

No. Flux models are run at cfg 1 in most workflows, which ignores the negative prompt. Describe what you want in the positive prompt instead.

Can I run Flux on a Mac?

Yes, ComfyUI runs on Apple Silicon Macs, but it is much slower than an NVIDIA GPU. Use the smaller Klein models or GGUF files and keep the resolution modest.

What is the difference between Flux dev and schnell?

Dev gives higher quality with 20 or more steps but has a non-commercial licence. Schnell is distilled for speed, needs about 4 steps and uses Apache 2.0.

Where do Flux text encoders go in ComfyUI?

Put clip_l and t5xxl files in ComfyUI/models/text_encoders/. The older models/clip folder also still works.

Sources

• ComfyUI docs: FLUX.1 text-to-image (files, folders, DualCLIPLoader)
• ComfyUI docs: FLUX.2 Dev
• ComfyUI docs: FLUX.2 Klein
• ComfyUI docs: FLUX.3 Klein (versions, steps, upsampler, ComfyUI 0.41.0+)
• Black Forest Labs on Hugging Face (model cards and licences)
• comfyanonymous/flux_text_encoders