flux2-prompt-engineering

FLUX.2 Prompt Engineering

Overview

FLUX.2 is Black Forest Labs' image generation and editing model family. It uses Mistral-3 24B VLM as its text encoder — it reads natural prose, not keyword lists. The single most important constraint: 30–80 words is the sweet spot. The model accepts up to 32K tokens but attention degrades beyond ~80 words, causing later instructions to be silently ignored.

When to Use

Use this skill when:

• Writing a FLUX.2 text-to-image prompt (any variant: max, pro, flex, klein)
• Editing an existing image with FLUX.2
• Reformulating a verbose, narrative, or poorly performing prompt
• Debugging FLUX output: flat lighting, default 3x3 grid layouts, miscounted items, ignored instructions, wrong colors, vague "professional photo" results
• Building a Replicate or BFL API payload for FLUX.2

Do NOT use this skill for:

• Stable Diffusion prompts (negative prompts work, keyword lists work, syntax differs)
• Midjourney prompts (--ar, --v, --style flags, weighting syntax differs)
• DALL-E / GPT-Image prompts (verbose narrative is preferred, not penalized)
• Imagen, Ideogram, or other non-FLUX models

The 12 rules below are FLUX.2-specific. Applying them to other models will degrade output.

Quick Reference Card

STRUCTURE:     Count + Surface → Per-item (position, description, color, angle) → Shadows → Camera → Size
WORD COUNT:    Target 60–80. Hard ceiling 100. Recount before sending — over runs are silent quality losses.
COLORS:        Hex codes per object: "wall #363636", "interior #B8B8B8"
CAMERA:        "Hasselblad X2D 80mm f/5.6" (never "professional photo")
LAYOUT:        Explicit positions: "upper left," "far lower right corner"
ANGLES:        Per-item: "straight" or "tilted N degrees" — alternate between them
SHADOWS:       Physical objects: "monstera shadow," "venetian blind stripes at 30°"
SURFACES:      "Micro-cement" for smooth modern; "concrete" only if rough is desired
TEXT:          In 'single quotes', keep 1–4 words, describe font style
NEGATIVE:      Never. Rephrase as positive description.
FORMATTING:    No em dashes, bullets, markdown. Continuous prose only.
COUNTS:        Exact number, never ranges ("six items" not "5–6 items")
UPSAMPLING:    On for T2I with short prompts. Off for edits.
EDITING:       One instruction per pass. Can't resize. Crop+upscale instead.
CROP TRICK:    10–15% inward crop + LANCZOS upscale = reliable zoom

The 12 Rules

1. No Negative Prompts

FLUX.2 does not support negative prompts. Negations often cause the model to generate exactly what you're avoiding.

BAD	GOOD
"no blur"	"sharp focus"
"no people"	"empty scene"
"no text inside frames"	"solid uniform cream interiors"
"photo areas contain no image"	"blank polaroid frames with solid gray interiors"
"no bright or saturated colors"	"muted warm neutral palette"

2. Exact Item Counts, Not Ranges

The model needs a single number to anchor its composition. Ranges cause ambiguity.

BAD	GOOD
"5–6 paper elements"	"Six items"
"several polaroids"	"Three polaroids"
"a few swatch cards"	"Two Pantone swatch cards"

3. Caveman Mode — Strip All Filler

Every word must carry visual information. The VLM processes meaning, not grammar. Articles, adjectives that don't change the visual output, and flowery descriptions waste tokens.

Filler patterns to cut:

• Preambles: "A top-down flatlay photograph of a designer's moodboard arrangement on..." → "Top-down flat lay, six items on..."
• Abstract aesthetics: "The overall aesthetic is minimal, editorial, sophisticated — like a high-end design studio still life" → "dark minimalist editorial"
• Redundant modifiers: "classic thick white borders (thicker at bottom)" → "white border"
• Hedging: "soft gray gradients catching the light bands" → "solid gray #B8B8B8 interior"
• Parenthetical clarifications: "(small, medium, large)" → assign sizes per item directly

4. Hex Colors Tied to Specific Objects

Always pair hex codes with the exact object. Abstract color references produce inconsistent results.

BAD	GOOD
"muted warm neutrals"	(assign per object below)
"anthracite/dark olive"	"anthracite #6B5F58"
"cool silver-gray"	"steel gray #8A9199"
"dark charcoal"	"#3A3A3A"
"cream-colored"	"off-white #E8E4DF"

5. Camera Model + Lens + Aperture

Naming specific equipment produces more authentic photorealism than abstract descriptors.

BAD	GOOD
"Shot with a medium format camera"	"Hasselblad X2D 80mm f/5.6"
"professional photo"	"Shot on Sony A7IV, 85mm, f/2.8"
"photorealistic"	(use camera spec instead)
"natural film-like grain"	"Kodak Portra 400 color science"

Focal length controls zoom level:

• 14–24mm: wide, dramatic, shows full scene
• 35–50mm: natural perspective
• 70–85mm: portrait/product, moderate crop
• 100mm+: telephoto, tight framing, items appear larger

6. Explicit Positions Per Item

Abstract layout descriptors ("scattered," "casual arrangement") are ignored — the model defaults to grid layouts. Assign every item a named position.

Position vocabulary: "upper left," "far upper right corner," "left middle edge," "center-right," "lower left," "far lower right corner," "bottom center near edge"

Each item gets its own sentence: position first, then description.

Upper left: large straight polaroid, white border, solid #B8B8B8 interior, washi tape top.

7. Per-Item Angles, Not Ranges

Angle ranges produce uniform tilt. Assign specific degrees to each item, and make some explicitly straight.

BAD	GOOD
"rotated at subtle casual angles (2–8°)"	Item A: "straight." Item B: "tilted 12 degrees." Item C: "tilted 6 degrees."
"each rotated slightly"	"Large items straight, small items tilted"

8. Name Physical Shadow Sources WITH Angles

"Soft shadows" or "warm natural light" produces flat lighting. Two requirements, both mandatory:

1. Name the physical object casting the shadow (monstera leaf, venetian blinds, leather portfolio, lamp, window frame).
2. Specify the angle ("at 30 degrees", "diagonal", or "horizontal" — never omit).

The word "soft" is a red flag — it almost always hides a missing physical source. Replace it.

BAD	GOOD
"warm directional sunlight"	"warm afternoon light from upper left"
"soft shadows from a window"	"venetian blind shadows at 30 degrees from window upper left"
"filtered through horizontal window blinds"	"diagonal venetian blind stripe shadows at 30 degrees"
"A monstera deliciosa leaf casts its organic shadow"	"monstera leaf shadow falling diagonally across upper left items"
"leather portfolio shadow across the corner"	"leather portfolio shadow at 45 degrees across lower right"

"Horizontal blinds" produces flat horizontal lines. "At 30 degrees" or "diagonal" produces realistic angled stripes. Both shadow types can coexist:

Monstera leaf shadow diagonally across upper left, venetian blind stripe shadows at 30 degrees across surface.

9. Surface Material Precision

"Concrete" produces heavily pitted, rough surfaces. Use interior finish terminology for modern aesthetics.

BAD	GOOD
"dark charcoal concrete wall with fine porous texture"	"smooth matte dark gray micro-cement wall #363636"
"concrete surface"	"micro-cement surface" (real finish term)

10. Text Rendering

FLUX.2 can render legible text. Use quotation marks around the text content and describe placement.

BAD	GOOD
reading MARCO LEDER / CREATIVE DIRECTOR	text reading 'MARCO LEDER' and 'CREATIVE DIRECTOR' in small sans-serif

Keep text short (1–4 words per line). Describe font style ("minimal sans-serif," "bold industrial lettering") and surface ("on the business card," "above the door").

11. No Formatting Characters in Prompts

The prompt is processed by a VLM text encoder, not rendered as markup. Formatting characters waste tokens and may confuse parsing.

Remove: em dashes (—), bullet points (•, -, *), markdown headers (#), parenthetical asides, line break formatting, ellipses (...)

Write as: continuous flowing prose, sentence after sentence.

12. Front-Load the Item Count

Start the prompt with the number of items and the surface. This anchors the model's composition.

Six items on smooth dark gray micro-cement wall #363636.

Prompt Structure

Priority order: Subject count + surface → Per-item descriptions → Lighting/shadows → Camera/technical → Dimensions

Word order matters — FLUX.2 weights earlier tokens more heavily. What comes first gets the most attention.

Length	Words	Use Case
Short	10–30	Quick concepts, style exploration
Medium	30–80	Sweet spot for most projects
Long	80–100	Complex multi-item scenes (hard ceiling)
Danger zone	100+	Attention degrades, later items ignored

Reformulation Workflow

When given a verbose or poorly structured prompt, follow these steps in order:

1. Count words. If over 80, the prompt needs cutting. Target 70–85 for complex scenes.
2. Eliminate negatives. Find every "no," "not," "without," "never," "none" and replace with positive descriptions.
3. Extract and count items. List every distinct visual item. Assign an exact count. Remove ranges and "several."
4. Assign positions and angles. Give each item an explicit position name and either "straight" or a specific degree of tilt. Alternate — not all items should be angled.
5. Convert colors to hex. Replace every color adjective with a hex code tied to the object.
6. Name shadow sources. Replace abstract lighting with physical objects: plant species, blind type + angle.
7. Specify camera. Replace "medium format camera" / "photorealistic" with model + lens + aperture.
8. Apply caveman compression. Cut articles (a, an, the), filler adjectives, preambles, aesthetic descriptions. Every surviving word must produce a visible difference.
9. Remove formatting characters. Strip em dashes, bullets, parentheticals. Continuous sentences only.
10. Reorder structure: Count + Surface → Items → Shadows → Camera → Dimensions.
11. Re-count words. If still over 85, cut further. If under 60 and using prompt_upsampling, that's fine — the upsampler will enrich it.

Pre-Flight Checklist

Before sending any prompt to FLUX.2, verify:

• Word count is 30–85
• Zero negative phrasing (no "no," "not," "without")
• Exact item count stated in first sentence
• Every item has an explicit position (upper left, lower right, etc.)
• Every item has an explicit angle (straight or N degrees)
• All colors are hex codes tied to specific objects
• Camera specified with model, lens mm, and f-number
• Every shadow has BOTH a named physical source AND an explicit angle ("at 30 degrees", "diagonal", or "horizontal")
• No "soft" hedges in lighting (it almost always means a missing physical source)
• Surface uses precise material term (micro-cement, not concrete)
• No em dashes, bullets, or markdown formatting in prompt text
• Text content in quotation marks with font style described
• No JSON inside Replicate payload (flatten to prose)
• prompt_upsampling is true for T2I, false for edits
• input_images omitted entirely for T2I (not empty array)

Full Before/After Example

BEFORE (~150 words, violates multiple rules)

A top-down flatlay photograph of a designer's moodboard on a dark charcoal concrete wall with fine porous texture. Warm directional sunlight enters from upper left, filtered through horizontal window blinds, casting soft diagonal bands across the scene. A monstera deliciosa leaf casts its organic shadow in the upper-left corner. Scattered across the wall are 5–6 matte paper elements, each rotated at subtle casual angles (2–8°): several blank Polaroid frames in varying sizes (small, medium, large) with classic thick white borders. The photo areas contain no image, just soft gray gradients. Two Pantone swatch cards — one anthracite/dark olive with PANTONE® logo, one with two blocks of dark charcoal and cool silver-gray. One cream-colored business card reading "MARCO LEDER / CREATIVE DIRECTOR / STUDIO" in minimal sans-serif. Color palette is muted warm neutrals. No bright or saturated colors. Shot with a medium format camera. Photorealistic.

Violations: 150 words (still ~2× over sweet spot), negative phrasing ("no image," "no bright colors"), range item count ("5–6"), abstract layout ("scattered"), angle range ("2–8°"), no hex codes, no camera specifics, em dashes, abstract shadow description, "concrete" surface.

AFTER (~90 words, all rules applied)

Top-down flat lay, six items on smooth matte dark gray micro-cement wall #363636. Upper left: tall Pantone card, white stock, PANTONE label, solid #B8B8B8 top block and anthracite #6B5F58 bottom block, washi tape, straight. Upper right: large straight polaroid, white border, solid #A8B0B8 interior, washi tape top. Center: medium polaroid tilted 8 degrees, solid #CCCCCC interior, washi tape. Lower left: Pantone card, two blocks charcoal #3A3A3A and steel gray #8A9199, labeled 'PANTONE ANTHRACITE' and 'STILL LIFE 03', tilted 12 degrees. Lower right: large straight polaroid, solid #B0B0B0 interior, washi tape. Bottom center: off-white #E8E4DF business card, rounded corners, text 'MARCO LEDER' and 'CREATIVE DIRECTOR' in small sans-serif, tilted 5 degrees. Monstera leaf shadow across upper left items, diagonal venetian blind shadows at 30 degrees across surface, warm afternoon light. Hasselblad X2D 80mm f/5.6, dark minimalist editorial. 2048x1324.

Use with: "prompt_upsampling": true — the dense structure provides layout; upsampling enriches material detail.

What Changed

Aspect	Before	After
Words	~150	~90
Item count	"5–6"	"six items"
Layout	"Scattered across"	Per-item positions
Angles	"2–8° range"	Per-item: "straight," "8 degrees," "12 degrees," "5 degrees"
Colors	"anthracite/dark olive"	"#6B5F58"
Negatives	"no image," "no bright colors"	"solid #B8B8B8 interior," removed
Camera	"medium format camera"	"Hasselblad X2D 80mm f/5.6"
Surface	"dark charcoal concrete wall"	"smooth matte dark gray micro-cement wall #363636"
Shadows	"filtered through horizontal window blinds"	"diagonal venetian blind shadows at 30 degrees"
Formatting	Em dashes, bullets	Continuous prose
Text	No quotes	'MARCO LEDER' in quotes with font described

Model Selection

Model	Best For	References
max	Highest fidelity, complex edits, character consistency	Up to 8 images
pro	Production at scale, good quality/cost balance	Up to 8 images
flex	Typography, fine-grained control (steps, guidance)	Up to 10 images
klein	Sub-second generation, real-time apps	Up to 5 images

API Parameters (Replicate)

{
  "width": 2048,
  "height": 1324,
  "prompt": "...",
  "resolution": "4 MP",
  "aspect_ratio": "custom",
  "output_format": "png",
  "output_quality": 80,
  "safety_tolerance": 2,
  "prompt_upsampling": true
}

input_images: Omit entirely for text-to-image. Do not send an empty array. For editing, provide the source image URL.

safety_tolerance: 0–6 scale. Default 2. Does not affect layout or quality.

JSON structured prompts: FLUX.2 supports JSON with fields like scene, subjects, camera. However, Replicate double-escapes inner JSON when embedded in the payload, corrupting the prompt. Always flatten JSON to natural language prose instead.

Prompt Upsampling

The prompt_upsampling parameter tells the model's text encoder to automatically expand your prompt with material, lighting, and atmospheric details.

USE when: Writing a dense caveman-mode prompt (60–85 words) for text-to-image generation. The dense prompt provides structure; upsampling fills in richness.

DO NOT USE when: Editing an existing image (rewrites edit instructions), or when your prompt is already 100+ words (pushes past attention limit).

Image Editing

What Works

Color swaps, style changes, adding/removing objects, background replacement, text changes.

What Does Not Work

Geometric resizing — "make cards 30% larger" has no meaning in diffusion space. The model thinks in semantics, not pixels.

Workaround: crop inward and upscale with LANCZOS.

from PIL import Image

img = Image.open("input.png")
w, h = img.size
pct = 15  # 10=subtle, 15=moderate, 20=aggressive
crop_x, crop_y = int(w * pct / 100), int(h * pct / 100)
cropped = img.crop((crop_x, crop_y, w - crop_x, h - crop_y))
result = cropped.resize((2048, 1324), Image.LANCZOS)
result.save("output.png")

Editing Rules

• One instruction per pass — multi-edit prompts cause selective execution
• Do NOT use prompt_upsampling for edits
• Do NOT overuse "same" — it anchors model to exact reproduction, preventing the requested change
• Describe the desired end state, not the transformation

Common Mistakes

Mistake	Symptom	Fix (Rule N)
Verbose 180-word prompt	Later instructions ignored, wrong items	Caveman compress to 80 words + prompt_upsampling: true (Rules 3, 12)
"Scattered," "casually arranged"	Default 3x3 grid layout	Per-item named positions (Rule 6)
Embedding JSON in Replicate payload	Corrupted prompt, garbage output	Flatten to natural language prose (API Parameters)
Saying "make cards 30% larger" in edit	No change to size	Crop + LANCZOS upscale (Image Editing)
Using "same" repeatedly in edit prompt	Requested change doesn't apply	Describe end state instead (Editing Rules)
"Concrete" wall	Heavily pitted, rough texture	"Micro-cement" for smooth modern (Rule 9)
"Soft natural light" / "warm sunlight"	Flat, lifeless lighting	Name the physical shadow source + angle (Rule 8)
"Multi-edit: change color and add a logo"	Only one edit applied	One instruction per pass (Editing Rules)
80mm focal length, want tighter crop	Subject too small	Bump focal length to 100–120mm (Rule 5)
Item count buried mid-sentence	Wrong number of items rendered	Front-load count + surface in first words (Rule 12)