FLUX.1 Dev: how to prompt it, best settings and creative ideas

Topic: Models

By Captain

Published 2026-09-30

A practical guide to FLUX.1 Dev by Black Forest Labs: natural-language prompting, guidance and step settings, LoRA stacking, text rendering and prompt ideas that play to its strengths.

Overview

FLUX.1 Dev

is a 12-billion-parameter rectified-flow transformer from Black Forest Labs. It reads prompts through a T5 text encoder, which means it understands full sentences, spatial relations and written text far better than the Stable Diffusion generation before it. It is guidance-distilled, so there is no

negative prompt

and the guidance value behaves differently from classic

CFG

The catalog entry is the fast fp8 Dev workflow: it accepts up to five

LoRAs

at once and finishes with a built-in 1.4x upscale. Dev is the model most Flux LoRAs are trained on, so it is the safest base for style, character and detail LoRAs.

Reference

Name	Type	What it is
Prompt	text	Plain sentences. Describe subject, setting, lighting, camera and mood in that order. No weights, no tag soup.
Guidance	1.5-5	3.5 is the default. Lower (2-2.5) gives softer, more photographic skin; higher (4-5) follows the prompt harder but flattens colors.
Steps	20-30	20 is enough for drafts, 28 for finals. Beyond 30 there is little gain.
Resolution	0.5-2 MP	1024x1024, 1216x832 or 1344x768 work natively; up to 2 MP without tiling.
LoRAs	up to 5	Style LoRAs at 0.6-0.9, character LoRAs at 0.8-1.0. Lower each weight when stacking three or more.

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

Write one paragraph that reads like a photo caption: who or what, where, what light, what lens or medium.
Keep guidance at 3.5 and steps at 24 for the first pass, then adjust guidance in 0.5 increments once the composition is right.
If you need legible text, put the exact words in double quotes inside the prompt and keep them short (one to four words).
Add LoRAs one at a time and check the result before adding the next; drop the weight of any LoRA that starts to bend anatomy.
Pick the best seed and rerun it with 28 steps for the final image.

Examples

Editorial portrait

A candid portrait of a silver-haired architect in her sixties standing in a half-finished concrete stairwell, dust in the air, soft window light from the left, shot on a 50mm lens with a slight film grain, muted teal and sand palette.

Text on a sign

A weathered wooden shop sign hanging over a cobblestone street in Lisbon reads "OPEN LATE" in hand-painted cream letters, late afternoon sun, tram cables overhead, documentary photography.

Creative idea: impossible product shot

A studio product photograph of a glass perfume bottle shaped like a thundercloud, tiny lightning glowing inside it, resting on wet black slate, single softbox from above, reflections crisp.

Tips

Flux has a recognisable default look: clean skin, centered subject, shallow focus. Mention film stocks, grain, or a specific camera to escape it.
Say what the camera is doing (low angle, wide lens, handheld) instead of listing adjectives; the T5 encoder follows directions.
Negative wording in the prompt (no hats) tends to add the thing. Describe what should be there instead.
For illustrations, name the medium (gouache, risograph, ballpoint pen) early in the prompt; it sets the whole rendering style.

Troubleshooting

Images look plastic and over-smooth

Why it happens

Guidance above 4 with portrait prompts pushes skin toward a retouched look.

How to fix it

Drop guidance to 2-2.5, add words like candid, natural skin texture, 35mm film, and keep steps at 24-28.

A LoRA has no visible effect

Why it happens

The LoRA was trained for Schnell or for another family, or its weight is too low under several other LoRAs.

How to fix it

Check the LoRA's base on its model page, raise its weight to 1.0 alone, then reintroduce the others at lower weights.

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.