How to Write FLUX.2 Prompts in ComfyUI: Step-by-Step Guide

If your FLUX.2 images look generic or don't match what you typed, the prompt is almost always the problem — not the model. FLUX.2 follows detailed instructions better than any previous Stable Diffusion model, but it reads prompts in a specific way. Paste an SD1.5-style prompt full of quality tags and you'll get a mediocre result. Write a clear description using the structure in this guide and FLUX.2 responds with precision.

This tutorial covers FLUX.2 prompting in ComfyUI from first principles — which node takes your prompt, how FLUX.2 weighs your words, and the exact techniques that produce professional results.

What You Need Before You Start

Models Needed

FLUX.2 uses a different set of model files from FLUX.1. You need four files — one diffusion model, two text encoder variants (pick fp8 for lower VRAM or bf16 for full precision), and one VAE. All are hosted on Hugging Face via the Comfy-Org repository.

Where to Place Each File in ComfyUI

Once you've downloaded the files, place them in the correct subfolders inside your ComfyUI installation. Note that FLUX.2 uses diffusion_models/ and text_encoders/ — not the unet/ and clip/ folders used by FLUX.1:

Not installed yet? Follow the how to install ComfyUI guide first, then come back here. You'll also need ComfyUI Manager for installing any missing nodes.

How Does FLUX.2 Read Your Prompt?

FLUX.2 was trained on a different architecture than SD1.5 or SDXL. It uses Mistral 3 Small as its text encoder — a compact but capable language model — which is why FLUX.2 can follow long, detailed, natural-language descriptions so precisely.

The side effect: FLUX.2 reads your prompt more like a human reads a sentence. What comes first matters most.

Word Order Is How FLUX.2 Decides What Matters

FLUX.2 weighs the first words of your prompt most heavily. Whatever you put at the start is treated as the primary focus of the image. Your subject must come first — not style tags, not quality words, not camera settings.

🔍 Click to zoom

FLUX.2 Uses the Mistral 3 Small Text Encoder

Unlike FLUX.1 which used T5-XXL + CLIP-L, FLUX.2 uses the Mistral 3 Small language model as its text encoder. The output feeds into a CLIP Text Encode node — which is where you type your prompt. You write one prompt. You don't split it between encoders.

🔍 Click to zoom

Quality Tags from SD1.5 Do Nothing Here

Words like masterpiece, best quality, and 8k are SD1.5 training conventions. FLUX.2 was not trained on them. They consume token weight on meaningless words and push your actual subject description further down the priority list.

The 4-Part Prompt Formula for FLUX.2

Every strong FLUX.2 prompt contains four components. You don't need to label them or follow a rigid order, but your prompt should cover all four before you generate.

leaning against a brick wall with arms folded, head tilted slightly downward
crouching beside a motorcycle, one hand resting on the fuel tank
ceramic bowl half-filled with miso soup, steam curling above the surface in still morning air

medium format film, Fuji Pro 400H, muted pastel tones
surveillance camera aesthetic, high contrast monochrome
architectural photography, tilt-shift lens, perspective correction

gouache illustration, thick paint strokes, visible texture
Soviet propaganda poster style, bold graphic shapes, limited palette
woodblock print, rough grain texture, earthy ink colours

single bare tungsten bulb overhead, deep shadows below
blue-hour dusk, city skyline silhouetted behind the subject
harsh midday sun from directly above, bleached concrete ground

good lighting
natural light
pretty light

🔍 Click to zoom

Where to Type Your Prompt in ComfyUI

This is the exact node setup for a FLUX.2 workflow. If you've loaded a working FLUX.2 workflow, your canvas already has these nodes. If not, load the FLUX.2 ComfyUI workflow first — download the JSON and drag it onto the ComfyUI canvas.

1. Find the node labelled CLIP Text Encode (Prompt) on your canvas — a rectangular block with a large multiline text field inside.
2. Click inside the text field. Type or paste your prompt directly here. There's no character limit to worry about for normal use.
3. Check that the output wire from this node (labelled CONDITIONING on the right side) connects to the positive input on your KSampler node.
4. Find the second CLIP Text Encode node connected to the negative input on the KSampler. For FLUX.2, leave this field completely empty — see the next section for why.
5. Click the orange Queue Prompt button in the top-right corner. A progress bar appears beneath it. Generation takes 15–60 seconds depending on your GPU.

You should see: a preview image appear in the output node on the right side of the canvas once generation completes.

🔍 Click to zoom

Why You Should Never Use Negative Prompts in FLUX.2

FLUX.2 does not support negative prompts the way SD1.5 or SDXL does. The model reads language literally — typing no blurry background can produce a blurry background. Typing no extra fingers can generate extra fingers.

This happens because FLUX.2's Mistral text encoder treats your negative prompt as a description, not an exclusion. It reads the word "fingers" and weights that concept in the output.

CFG Scale — Set This Before Generating

FLUX.2 runs best at CFG 1.0–3.5. The high CFG values (7–15) that work for SD1.5 will distort FLUX.2 images — faces melt, colors saturate unnaturally, and details become overcooked. Find the KSampler node on your canvas and set cfg to 1.0 as a starting point.

Example Prompts — Ready to Use in ComfyUI

These four prompts use JSON structured format and cover a range of scenes — documentary street photography, noir atmosphere, editorial travel, and gritty urban realism. Copy any prompt directly into your CLIP Text Encode node and generate.

📍 Kolkata, India

Monsoon Chai Stall

Copy Prompt▶View Prompt

📍 Yurakucho, Tokyo

Tokyo Yakitori Alley — Late Night

Copy Prompt▶View Prompt

📍 Medina of Marrakech

Marrakech Spice Souk — Golden Hour

Copy Prompt▶View Prompt

📍 Lower East Side, Manhattan

New York Bodega — 3am

Copy Prompt▶View Prompt

Advanced Prompting Techniques for FLUX.2

Once your basic workflow is producing clean results, these three techniques give you precise control over specific output characteristics.

How to Control Exact Colors Using HEX Codes

FLUX.2 understands HEX color codes — the six-character codes used in web design to specify exact colors (e.g. #C0392B for a specific shade of brick red). Pair the HEX code directly with the object it applies to.

How to Add Text Inside Your Images

FLUX.2 renders text significantly better than SD1.5 or SDXL. To generate readable text in an image, wrap the exact wording in quotation marks inside your prompt.

Specify placement and style alongside the text:

bold condensed type reading "SOLD OUT" stamped in red ink diagonally across a white paper label
hand-painted text reading "fresh bread daily" in black cursive on a chalkboard menu
a door plate reading "ROOM 7" in brass relief lettering, close-up macro shot

How to Use JSON Structured Prompting

For complex scenes with multiple subjects or many independent variables, JSON format gives you modular control. Each element is a separate field — you can change the camera angle without rewriting the entire prompt. Paste this directly into the CLIP Text Encode node:

🔍 Click to zoom

Prompt Length — How Long Should Your FLUX.2 Prompt Be?

FLUX.2 supports prompts up to 32,000 tokens — far more than you'll ever need. Length is not the goal. Precision is. The practical rule: every word you add should visibly change something in the output. If you remove a word and the image doesn't change, that word was padding.

Build Prompts in Stages

This approach also makes it easier to identify which part of your prompt caused an unexpected result.

Stage 1 (subject only):
An elderly woman selling dried spices at an open-air market stall

Stage 2 (add location):
An elderly woman selling dried spices at an open-air market stall, wooden trestle table covered with hessian sacks and terracotta bowls filled with turmeric, cumin, and paprika

Stage 3 (add style and lighting):
An elderly woman selling dried spices at an open-air market stall, wooden trestle table covered with hessian sacks and terracotta bowls, travel documentary photography, 50mm lens, bright diffused midday shade under a canvas awning

Stage 4 (add final detail):
An elderly woman selling dried spices at an open-air market stall, wooden trestle table covered with hessian sacks and terracotta bowls filled with turmeric, cumin, and paprika, travel documentary photography, 50mm lens, bright diffused midday shade under a canvas awning, shallow depth of field pulling focus to her hands scooping spice, warm ochre and brick-red colour palette, slight film grain

Troubleshooting — Why Your FLUX.2 Prompt Isn't Working

The Image Looks Generic No Matter What I Type

Cause: The prompt is too short or too vague. FLUX.2 fills in missing information — and the default fill is often stock-photo generic.

Fix:

1. Add subject specifics — not a man but a man in his mid-40s, closely cropped beard, steel-rimmed glasses, wearing a worn grey henley, relaxed posture.
2. Add a lighting description — single bare bulb overhead casting a hard downward shadow changes the feel more than most other additions.
3. Add a style reference — Magnum-style documentary photography or medium format film, muted tones gives FLUX.2 a visual language to work in.

The Image Has Distorted Hands or Extra Fingers

Cause: Hands at complex angles are architecturally difficult for any diffusion model. Using negative prompts to fix this makes it worse — FLUX.2 reads no extra fingers as a description containing the concept "extra fingers."

Fix:

1. Move hands out of the composition: hands out of frame, waist-up shot.
2. Describe a specific hand position if hands must appear: left hand open flat on the table, four fingers together, thumb extended.
3. Generate multiple times and select the cleanest result — iteration is the most reliable fix.

My Colors Look Wrong or Change Between Generations

Cause: Color language like "bright blue" or "warm tones" is interpreted loosely and changes across seeds (a seed is a number that controls which random variation ComfyUI generates — find it in the KSampler node).

Fix:

1. Use HEX codes tied directly to the object: a linen shirt in dusty sage green #8FAF8F.
2. Name the light source explicitly — sodium vapour street lamp produces orange-shifted images; cloudy north-facing window light produces cool, flat-lit images.
3. Lock your seed in the KSampler node when testing color changes so you isolate the prompt variable.

FLUX.2 Is Ignoring Part of My Prompt

Cause: Critical details are buried at the end of a long prompt. FLUX.2 applies more weight to the beginning of the prompt.

Fix:

1. Move your most important element to the first sentence.
2. Check whether your style tag appears after 60+ words of scene description — if so, move it to sentence two.
3. Shorten the prompt. A focused 40-word prompt usually beats a cluttered 120-word prompt.

Frequently Asked Questions

Published: 2025-06-15 · Last updated: 2025-06-15 · Tested on RTX 3060 (12 GB VRAM) and RTX 4080 (16 GB VRAM) · ComfyUI v0.3.10