Long or stylised string, guidance too high. Shorter quoted text, plain font, guidance 4, several seeds.

Role not stated in the prompt. Name each image and what to take from it.

High guidance and no lens / film words. Guidance 3, add film grain and lens, use a realism or skin LoRA from the family page.

fp16 weights or 4 MP on a 16 GB card. fp8 or GGUF, 1 MP drafts, offload; or hosted.

the best open source ai models

FLUX.2 prompting guide: natural-language prompts, multi-reference, editing and LoRAs

Q: FLUX.1 LoRA gives errors or nothing

Different architecture. Use LoRAs from the FLUX.2 family only; the catalog separates them.

Topic: Prompting

By Captain

Published 2026-05-28

Prompting FLUX.2 dev for the results it is known for: paragraph prompts with subject, scene, light and lens, exact text in quotes, up to several reference images for products, people and styles, instruction-style edits, the guidance and step settings that matter, and how FLUX.2 LoRAs differ from Flux dev LoRAs.

Overview

FLUX.2

is Black Forest Labs' second-generation image model: a single network that generates from text, edits an existing picture from an instruction and composes from several reference images, at up to 4 megapixels, with text rendering far ahead of

FLUX.1

. The open dev weights made it the fastest-growing

LoRA

family on this site within weeks.

The model reads long prompts as sentences and is sensitive to structure: who or what, where, the light, the lens or medium, then details. Tag soup and (weights:1.3) do nothing. Exact text goes in quotes. For editing you give one or more images and say what to change; for references you describe how each image should be used.

FLUX.2 dev

is large (fp8 on 24 GB is comfortable, GGUF Q4 on 16 GB works with offload).

PirateDiffusion

and

BitVector

host it with the popular LoRAs installed, and every FLUX.2 model page on this site has a Run box for both.

Reference

Name	Type	What it is
Prompt structure	text	Subject and action, setting, lighting, camera or medium, colour and mood, details. One to three paragraphs are fine.
Quoted text	text	Words in double quotes are rendered as written: a neon sign that reads "OPEN LATE".
Reference images	input	Several pictures; name their roles: "the person from image 1 wearing the dress from image 2 in the style of image 3".
Edit instruction	text	With one input image: "Replace the sky with a thunderstorm, keep the building unchanged".
Guidance	setting	3 to 4 for photos, 4 to 5 for graphic design and text, lower (2 to 3) for painterly looks.
Steps	setting	28 to 40; distilled / turbo LoRAs bring it to 8.
Resolution	setting	Any aspect ratio from 1 to 4 MP; 1024x1024 for drafts, 2048 on a side for finals.
LoRA weight	setting	0.7 to 1.0. FLUX.2 LoRAs are not compatible with FLUX.1 ones.
Text encoder	file	Mistral-based encoder shipped with FLUX.2 (fp8 version for 16 GB cards).

No install required - run AI models on the cloud

BitVector Prism is the easiest way to start: pick a model, type a prompt and generate in a clean web app, with nothing to configure. BitVector is also available on Discord and on the web (SpyGlass).

Try BitVector

Step by step

Load the FLUX.2 dev template in
ComfyUI
(Browse Templates > Image > Flux 2), point the loaders at the diffusion model (fp8 or GGUF), the text encoder and the VAE.
Write a sentence prompt in the order subject, setting, light, lens, details. Start at 1024x1024, 30 steps, guidance 3.5.
Add exact text in quotes if the image needs it; keep each string short and name the surface it sits on.
For references, add Load Image nodes into the reference inputs and describe each image's role in the prompt.
For edits, feed one image and give an instruction; say what must stay.
Add a LoRA from the FLUX.2 family page (realism, skin, product, anime), weight 0.8, trigger word first in the prompt.
Finals at 2 MP or more; use the turbo LoRA if time matters.
Hosted: the Run box on any FLUX.2 model page sends it to PirateDiffusion (hosted workflow) or BitVector (web app).

Examples

Photo

A woman in her sixties with short grey hair laughs while pouring coffee in a sunlit kitchen, morning light through linen curtains, 50mm lens, shallow depth of field, warm film tones, natural skin texture

Design with text

A minimalist poster for a jazz night, large white serif title "BLUE HOUR" centered, subtitle "Friday 9 PM" below, deep navy background with a single gold saxophone silhouette, lots of negative space

Multi-reference

The man from image 1 wearing the leather jacket from image 2, standing in the rainy street of image 3 at night, neon reflections, cinematic

Edit

Replace the overcast sky with a dramatic sunset, keep the lighthouse and the rocks exactly as they are, adjust the light on the scene to match

Tips

Lens and film words steer realism strongly: "85mm, f/1.8, Kodak Portra 400" or "wide 24mm, deep focus, harsh noon sun".
Describe light before you describe colour; FLUX.2 builds the scene from light.
For product shots, a clean reference image plus "keep the product exactly as in image 1" beats any description.
Avoid negatives ("no people"); say what is there instead ("an empty street").
Guidance above 5 saturates colours and burns text; raise steps instead.
FLUX.2 LoRAs trained on 20 to 40 images at 1024 give strong identity; the family page shows which ones the community keeps downloading.
The weekly ranking is the quickest way to see new FLUX.2 LoRAs; the family had over a thousand within its first season.

Use LoRAs from the FLUX.2 family only; the catalog separates them.

No install required - run AI models on the cloud

PirateDiffusion is Telegram only and built for pros: thousands of models, LoRAs and workflows driven by chat commands, with unlimited generation on a fixed price plan.