AI Folks
  • Courses
  • Web Of AI
  • Pricing
  • Enterprise
Log in
PreviousYour first ComfyUI workflowNextEditing and control

Vibe Coder · Unit 20: Image Models: Flux, ComfyUI and LoRAs · Lesson 4 of 7

Flux in ComfyUI

Swap the small starter model for Flux: start from a template, put three model files in the right folders, and feed it reference images.

Read the short notes and pictures.Do everything inside a blue frame.

From a sketch to a real model

The seven-node workflow used a small starter model so you could see the parts. Flux is far more capable, and it is put together a little differently: its three parts arrive as separate files instead of one checkpoint. You do not need to wire that by hand.

Start from a template

  1. 1

    Open the Workflow Templates

    ComfyUI ships ready-made workflows for current models, including FLUX.2.

  2. 2

    Pick the one that matches your job

    Text to image, image editing, or generation with reference images.

  3. 3

    Let it tell you what is missing

    A template lists the model files it needs. Download each into the folder it names.

  4. 4

    If a template is not there

    Your ComfyUI may be out of date. Update it. On Comfy Cloud the templates open directly.

Three files, three folders

ComfyUI’s FLUX.2 [dev] guide names three files and where each belongs. The same layout applies to the klein models with their own file names.

The files

flux2_dev_fp8mixed.safetensors
mistral_3_small_flux2_bf16.safetensors
flux2-vae.safetensors

Where they go

ComfyUI/models/diffusion_models/
ComfyUI/models/text_encoders/
ComfyUI/models/vae/
The diffusion model removes the noise, the text encoder reads your prompt, the VAE turns the latent into pixels. A LoRA, which you will train later, goes in models/loras/.
A wooden workbench, hand-drawn. A table lists three files and the folder each belongs in.
The diffusion model, the text encoder and the VAE each have a folder.

Size and speed

The dev files in the guide are quantized, which means they are stored at lower precision to use less memory. They are still large downloads. If your machine struggles, use a klein model, which is built for consumer graphics cards, or run the template on Comfy Cloud.

Writing the prompt

Current Flux models have a strong text encoder, so they follow full sentences. Write a description, not a pile of keywords.

Vague

invoice app, modern, cool, 4k

Specific

A laptop on a pale oak desk showing an invoicing dashboard. Morning light from the left. A ceramic mug beside it. Clean editorial product photography. The screen header reads "Tally".
The second prompt names the subject, the setting, the light, the style and the exact text. Each of those is something you can change on its own.

Reference images

A product has to look like the same product in every picture. FLUX.2 accepts reference images for this: you load your product shot or your style sample through a Load Image node and the model keeps it consistent while changing the scene. Klein takes up to 4 reference images.

A wooden workbench, hand-drawn. Two Load Image nodes feed a FLUX.2 node, which feeds Save Image.
Reference images go in through Load Image nodes.
ActivityYour turnWrite it yourself first, then compare.

You need the same mug-and-laptop scene in three versions: morning, evening, and with a plant added. What do you keep fixed and what do you change for each?

ActivityBuild stepA brand image kit

Make three images with Flux

Load a FLUX.2 template, put the model files in place (or use Comfy Cloud), and make three images from your brief. Use one reference image so the product or character matches across them. Save the workflow with a clear name.

Done when

Keep the saved workflow file. It is part of what you publish at the end of the unit.

Keep these

  • Start from a workflow template, not an empty canvas.
  • Flux is three files: diffusion model, text encoder and VAE.
  • Reference images keep a product or character consistent.
Previous lessonYour first ComfyUI workflowNext lessonEditing and control