flux2dev-prompt-engineering-master

SKILL: FLUX.2 Dev Prompt Engineering Master

Version: 1.0 | Hermes Agent Universal Flux2 Prompt Optimization Doctrine

---

DESCRIPTION

The master reference for FLUX.2 Dev prompt engineering across all domains. This skill distills the deep research on FLUX.2 Dev's 32B guidance-distilled architecture, token-weight hierarchy, optimal parameter configurations, JSON structured prompting, and the positive-only constraint system into a single authoritative reference. Use this skill as the foundation before invoking any domain-specific FLUX.2 Dev skill. Covers prompt length optimization, camera specification syntax, color control, text-in-image rendering, multi-language support, and the complete anti-pattern catalog.

---

TECHNICAL PARAMETERS

FLUX.2 Dev Model Architecture (What Makes It Different):

• Parameters: 32 billion (vs. FLUX.1's 12B) — significantly higher capacity for detail adherence
• Architecture: Rectified flow transformer + Mistral-3's 24B vision-language model
• Guidance: Distilled — runs guidance scale 3–5 natively (not 7–12 like SDXL)
• Negatives: No native negative prompt support — all suppression must be positive-framed
• Prompt Encoding: Dual encoder (T5 + CLIP) — natural language descriptions outperform keyword lists
• Token Limit: T5 encoder supports 512 tokens; FLUX.2 Dev summarizes prompts >200 words internally
• Weight Syntax: Does NOT support (word:1.2), ++, --, or any bracket-weight syntax
• Resolution: Native up to 4MP (Pro), 1024–1536 base recommended for Dev
• Text Rendering: Best-in-class open model — quotes, font specification, placement all functional
• HEX Support: Direct hex color codes followed with high fidelity when tied to elements
• Multi-Language: Native multilingual understanding without translation

The Token-Weight Hierarchy (Critical): FLUX.2 Dev applies extreme early-token weighting. The first 10–15 tokens determine 60%+ of the output character.

Token Position	Weight	What to Place	Example
1–5	Maximum (≈35%)	Subject identity + scale	"Titanium smartwatch" / "Wide establishing shot"
6–10	Very High (≈20%)	Key descriptor / action	"floating at angle" / "lone rider on horseback"
11–20	High (≈20%)	Secondary subject detail	"matte black ceramic bezel" / "crossing desert plain"
21–40	Moderate (≈15%)	Environment / context	"against clean neutral gray" / "at golden hour"
41–70	Low (≈7%)	Technical / camera	"Shot on Phase One IQ4" / "85mm at f/1.4"
71+	Minimal (≈3%)	Atmosphere / mood	"melancholic twilight atmosphere"

Optimal Prompt Length by Use Case:

Length	Word Count	Use Case	Risk
Ultra-Short	< 15 words	Style exploration, quick concepts	Model hallucinates extensively
Short	15–30 words	Drafts, mood boards, rapid iteration	Limited control over specifics
Sweet Spot	30–80 words	Most production work, ideal balance	Optimal control with no summarization
Long	80–120 words	Complex scenes, multiple subjects	Minor summarization risk
Too Long	120–200 words	Detailed architectural, multi-character	Internal summarization drops elements
Excessive	200+ words	Everything specified	FLUX.2 compresses heavily; details lost

Guidance Scale Calibration:

Scale Range	Effect	Best For	Risk
2.5–3.0	High creative freedom, loose adherence	Artistic exploration, happy accidents	Subject drift, ignored details
3.0–3.5	Balanced creative freedom	Mood shots, atmospheric work	Minor detail loss acceptable
3.5–4.0	Recommended default for most work	Cinematic stills, portraits, editorial	Sweet spot for most domains
4.0–4.5	Strong adherence, slight saturation	Product photography, architecture, technical	Slight over-saturation possible
4.5–5.0	Maximum adherence, highest saturation	Brand assets, HEX accuracy critical	Risk of posterization, plastic skin
5.0–7.0	Extreme adherence, over-processed	NOT RECOMMENDED for FLUX.2 Dev	Over-saturated, artifact-heavy
7.0+	Broken — do not use	Never	Severe over-saturation, doubled elements

Step Count Optimization:

Steps	Quality	Speed	Recommended Use
16–20	Draft quality, soft detail	Very fast	Concept exploration, composition testing
20–24	Acceptable, some softness	Fast	Rapid iteration, style matching
24–28	Good quality, adequate detail	Moderate	Quick production, social media
28–35	Production quality, crisp detail	Standard	Final assets, most professional work
35–40	Maximum detail, grain fidelity	Slower	Macro, texture, technical documentation
40–50	Diminishing returns	Slowest	Only when micro-detail is paramount

---

THE POSITIVE-ONLY CONSTRAINT SYSTEM

FLUX.2 Dev does not support negative prompts natively. Every exclusion must be reframed as a desired positive quality.

Universal Positive Reframe Dictionary:

Problem Area	Negative (Never Use)	Positive Reframe (Always Use)
Blur	not blurry, no blur	sharp focus, crisp detail, tack-sharp
Low quality	not low quality, no low quality	high fidelity, detailed rendering, precise execution
Anatomy errors	no extra fingers, bad anatomy	natural human anatomy, correct digit count, balanced proportions
Text/watermark	no text, no watermark	clean image, no visible text, pure visual content
Plastic skin	not plastic, no AI look	natural skin texture, visible pores, organic surface
Over-saturation	not oversaturated	natural color balance, realistic saturation, true-to-life tones
Artifacts	no artifacts, no distortion	clean rendering, coherent geometry, stable composition
White background issues	not white background	clean neutral backdrop, minimal environment, light gray seamless
Floating objects	not floating	grounded on surface, natural contact shadow, weight and mass
Conflicting styles	not mixed styles	cohesive visual approach, unified aesthetic, consistent style

The Guardrails Clause (Append to Every Prompt):

natural proportions, coherent spatial logic, realistic light behavior,
authentic material texture, stable composition, accurate perspective,
no artificial smoothing, no impossible geometry, no conflicting light sources

---

PROMPT ARCHITECTURE

The Universal FLUX.2 Dev Template:

[Subject + identity] — FIRST 5 TOKENS, MAXIMUM WEIGHT
[Action / pose / state] — Tokens 6–10, very high weight
[Key detail / feature] — Tokens 11–20, high weight
[Environment / setting] — Tokens 21–40, moderate weight
[Camera + lens + aperture] — Tokens 41–60, moderate weight
[Lighting: source + direction + quality + color temp] — Tokens 61–80, low weight
[Style / film stock / era] — Tokens 81–100, low weight
[Atmosphere / effects / grain] — Tokens 101–120, minimal weight
[Guardrails clause] — Final tokens, minimal but present

JSON Structured Prompt Template (For Complex Scenes):

{
  "scene": "Overall environment description",
  "subjects": [
    {
      "type": "subject_category",
      "description": "Detailed subject description",
      "position": "foreground/midground/background",
      "color": "#HEXCODE"
    }
  ],
  "style": "Photographic or artistic approach",
  "colors": {
    "palette": ["#HEX1", "#HEX2", "#HEX3"],
    "dominant_hue": "Descriptive color name",
    "color_temperature": "XK"
  },
  "lighting": {
    "key_light": "Source, direction, quality",
    "fill_ratio": "Key:Fill ratio",
    "color_temp_k": "NumberK",
    "practical_sources": "Visible light sources in scene"
  },
  "camera": {
    "lens": "Focal length and type",
    "aperture": "f-stop",
    "shot_type": "Wide/medium/close-up/etc.",
    "angle": "Eye-level/low/high/Dutch"
  },
  "atmosphere": {
    "haze": "None/volumetric/light",
    "grain": "Fine/medium/heavy",
    "effects": "Lens flare, chromatic aberration, vignette"
  }
}

---

ADVANCED TECHNIQUES

1. The Camera-First Photorealism Hack

• Principle: Camera specifications trigger FLUX.2's photorealistic rendering pipeline more reliably than quality tokens.
• Technique: Always include "Shot on [specific camera], [lens] at f/[aperture]" in the first half of the prompt.
• Comparison:
  • Weak: "Photorealistic portrait of a woman, high quality, detailed"
  • Strong: "Portrait of a woman, shot on Canon EOS R5 with 85mm f/1.2 at f/2.0, detailed"
• Why it works: FLUX.2 was trained on image-caption pairs where camera EXIF data frequently appeared. Camera names are strong conditioning tokens.

2. The Behavioral Light Description

• Principle: Describing what light DOES outperforms naming what light IS.
• Technique: Replace lighting names with behavioral descriptions.
• Comparison:
  • Weak: "Golden hour lighting, warm tones"
  • Strong: "Warm golden sunset light streaming through tall windows, casting long shadows across hardwood floor, dust particles visible in light beam"
• Why it works: FLUX.2's VLM backbone processes scene descriptions as spatial arrangements. Behavioral light descriptions encode spatial relationships.

3. The HEX Color Lock

• Principle: FLUX.2 follows HEX codes with high fidelity when properly formatted.
• Technique: Tie HEX to specific elements, not used standalone.
• Comparison:
  • Weak: "#FF6B35 and #004E89"
  • Strong: "Running shoe with primary upper in exact color #FF6B35, accent swoosh in #004E89, white midsole"
• Why it works: HEX codes without element binding are interpreted as general color mood. Element binding creates precise color assignment.

4. The Text-in-Image Precision Method

• Principle: FLUX.2 has best-in-class text rendering among open models.
• Technique: Use quotation marks, specify placement, define typography.
• Rules:
  • Always quote exact text: "The sign reads 'OPEN LATE'"
  • Specify font separately: "in bold sans-serif font"
  • ALL CAPS in prompt = ALL CAPS in image
  • Shorter text = more accurate: 2–5 words most reliable
  • Placement: "At top of poster," "Centered on banner"
  • Clean background behind text improves accuracy significantly

5. The Three-Pass Iteration Protocol

• Pass 1 — Foundation: Basic prompt with core subject, camera, and light. Generate at 20 steps, guidance 3.5.
• Pass 2 — Refinement: Analyze output. Add missing details. Adjust light direction. Generate at 28 steps, guidance 4.0.
• Pass 3 — Polish: Fine-tune atmosphere, add grain/effects, perfect color. Generate at 35 steps, guidance 4.0–4.5.
• Cost efficiency: Pass 1 uses fast generation. Only Pass 3 uses maximum quality settings.

6. The Multi-Language Native Prompt

• Principle: FLUX.2 understands prompts in multiple languages without translation.
• Technique: Write naturally in your native language maintaining the same structure.
• Examples work in: French, German, Spanish, Italian, Portuguese, Dutch, Russian, Chinese, Japanese, Korean, Thai, Arabic, Hindi
• Structure remains: Subject → Action → Style → Context regardless of language
• Cultural references function natively: "Seoul rooftop garden at sunset" works in Korean with identical visual output

7. The Temporal Consistency Chain

• Principle: When generating multiple related images (storyboard, campaign, lookbook), consistency tokens lock visual identity.
• Technique: Include a scene identifier and locked parameters in every prompt:

Scene 3B — Interior Night. Camera locked: ARRI Alexa 65, Cooke S4/i 50mm, Kodak Vision3 500T.
Subject: [varies per shot]. Light: [varies within 200K]. Atmosphere: [consistent haze level].

• Why it works: FLUX.2 processes scene identifiers as contextual anchors. Repeating locked parameters reinforces conditioning.

---

ANTI-PATTERN CATALOG

Forbidden Patterns in FLUX.2 Dev Prompting:

Anti-Pattern	Why It Fails	The Fix
"Cinematic" standalone	Too vague; FLUX.2 defaults to generic movie still	Replace with specific film stock + lens + aspect ratio
"Beautiful," "stunning," "amazing"	Quality judgment words carry zero visual information	Replace with specific technical parameters
"White background"	Causes fuzziness and quality degradation in Dev	Use "clean neutral studio environment" or "light gray seamless"
Keyword soup	"Woman, dress, red, beautiful, high quality, 8k" — FLUX.2 processes natural language, not keyword lists	Write in complete descriptive sentences
Midjourney parameter flags	--ar 16:9, --v 6, --no text are ignored and waste tokens	Describe composition: "2.39:1 aspect ratio composition"
Prompt weights (word:1.2)	FLUX.2 does not support weight syntax; may confuse model	Use natural emphasis: "prominently featuring," "especially detailed"
Conflicting light descriptions	"Bright midday sun AND moody dark shadows" — FLUX.2 averages into muddy neutral	Pick ONE lighting scenario; describe it thoroughly
Burying subject at end	"In a beautiful garden with flowers and sunlight, a woman sits" — subject gets minimal weight	Lead with subject: "A woman sits in a garden..."
Abstract style without anchor	"Make it look professional" provides no visual instruction	Anchor to specific camera: "Shot on Phase One IQ4, Schneider 80mm"
Negatives in positive-only model	"No extra fingers" can paradoxically trigger finger artifacts	Reframe: "Natural hand anatomy with correct five-finger count"
Overloading >200 words	FLUX.2 summarizes long prompts; details get compressed or dropped	Keep to 30–80 words; use JSON for complex multi-element scenes
Generic quality tokens	"8K, ultra HD, highly detailed, masterpiece" — FLUX.2 ignores these	Replace with specific detail descriptors: "Individual hair strands visible"

---

EXAMPLE PROMPTS

Weak vs. Strong Comparison (Portrait):

❌ Weak:

Beautiful portrait of a woman, high quality, detailed, professional photography, 8k, stunning

✅ Strong:

Portrait of a 35-year-old woman with auburn hair, shot on Canon EOS R5 with 85mm f/1.2 at f/2.0, soft window light from camera-left creating gentle Rembrandt pattern, visible skin texture with natural pores and subtle freckles, warm 4000K light, shallow depth of field with cream-colored bokeh background, individual hair strands catching light, sharp focus on near eye with visible catchlight, natural human anatomy, no artificial smoothing

Weak vs. Strong Comparison (Product):

❌ Weak:

A black coffee mug on white background, high quality, product photo, professional, detailed

✅ Strong:

Matte black ceramic coffee mug with visible throwing rings and slight irregular rim, steam rising from hot coffee surface, sitting on polished concrete countertop with subtle reflection, three-point softbox lighting with diffused key from left and soft fill from right, shot on Phase One IQ4 with 120mm macro at f/5.6, clean neutral studio environment with soft gradient, sharp focus on mug surface and steam texture, commercial product photography, natural shadow grounding the mug

Weak vs. Strong Comparison (Architecture):

❌ Weak:

Beautiful modern house on hill, stunning architecture, high quality, amazing view, professional

✅ Strong:

Modernist concrete villa cantilevered over coastal cliff, board-formed concrete texture with visible timber grain pattern, floor-to-ceiling glass curtain wall reflecting ocean and sky, warm 3000K golden hour light from west illuminating concrete facade, shot on technical camera with 24mm tilt-shift at f/11 maintaining vertical lines, two-point perspective from approach path, native coastal vegetation in foreground, deep blue Pacific Ocean in background, Kodak Vision3 250T film stock with fine grain, architectural photography, awe-inspiring coastal integration

---

TECHNICAL NOTES FOR AI GENERATION

• This skill is the foundation: Always consult this skill before any domain-specific FLUX.2 Dev skill. The principles here govern all domain-specific applications.
• Subject-first is non-negotiable: The first 10 tokens determine the entire output character. Never bury the subject.
• Camera specs > quality tokens: "Shot on Canon EOS R5" triggers better photorealism than "photorealistic, 8K, highly detailed" combined.
• Behavioral descriptions > naming: Describe what light DOES, what fabric DOES, what shadow DOES — not what they ARE.
• JSON for complexity: When a scene has 3+ subjects, 2+ light sources, or brand-critical HEX colors, use JSON structured prompts.
• Guidance 3.5–4.5 is the safe zone: Below 3.0 = drift. Above 5.0 = over-processing artifacts.
• Steps 28–35 for production: Below 25 = soft. Above 40 = diminishing returns with increased artifact risk.
• Always append guardrails: The positive-only guardrails clause should end every prompt as a safety net.
• Iterate in three passes: Foundation (fast/cheap) → Refinement → Polish (slow/expensive).
• For batch consistency: Use identical locked parameters across all prompts in a series. Vary only the intended variable.
• Multi-language is native: No translation needed. FLUX.2 processes the semantic content regardless of language.
• Text rendering: Quotes + placement + font + short length = highest accuracy for in-image text.
• Never use "white background" in Dev: This is the single most damaging phrase for FLUX.2 Dev output quality. Always reframe.
• When in doubt, be specific: "Red" → "Crimson silk with slight orange undertone." "Tall building" → "42-story glass tower with setbacks at 15th and 30th floors."