🎨 Sepia Be Gone

A portable prompt skill to banish the yellow/orange/sepia filter from AI-generated images.
Works with opencode, Claude Code, Codex, Cursor, Windsurf, VS Code Copilot, and any image generator.

"AI doesn't have to look like it was developed in a 1970s darkroom."

---

🤔 The Problem: Why AI Images Look Yellow

AI image models don't just "learn from old paintings." The yellow/orange cast comes from token bias in training data:

🎯 Prompt Keyword	📊 Statistical Association	🎨 Result
cinematic	Orange/teal grading (Hollywood LUTs)	Heavy amber cast
golden hour	Warm sunset lighting	Extreme yellow/orange
premium, luxury	Warm "golden" commercial photography	Cream whites, brown blacks
appetizing (food)	Golden food photography overlays	Greens→brown, whites→cream
studio lighting	Often tungsten (3200K) not daylight	Warm color temperature
dramatic lighting	Chiaroscuro with warm key lights	Orange shadows

RLHF amplifies this: Human raters consistently prefer warmer images for "appeal," so models learn warmth = quality.

---

✨ What This Skill Does

✅ Preserves	🎯 Applies	🚫 Removes
Layout & composition	Neutral 5600K daylight balance	Yellow tint
Products & packaging	Clean whites (RGB 255,255,255)	Orange cast
Logos & labels	Deep neutral blacks (RGB 0,0,0)	Warm filter
Prices & addresses	Accurate product/brand colors	Sepia tone
ALL visible text	Natural contrast, crisp lighting	Amber / golden hour
Design structure & spacing	Modern commercial look	Vintage grading

---

🖼 Before → After Examples — See It In Action

Real results from the skill — no cherry-picking, just the prompt applied to yellow-tinted images.

---

1️⃣ 🛒 Supermarket Promo Poster

Problem: Yellow/orange cast — cream whites, amber shadows, inaccurate brand blues

Before (Yellow Tint)	After (Neutral)

Preserved: Price ($3.99), address (123 Main St), product placement, logo, layout
Fixed: 5600K daylight balance, clean whites, neutral grays, accurate brand blues
Prompt used: Full prompt from prompts/neutral_color_balance.md

---

2️⃣ 🍔 Food Photography

Problem: Golden "appetizing" overlay — brownish greens, yellow whites, artificial warmth

Before (Warm Filter)	After (Color-Accurate)

Preserved: Food appeal, composition, plating
Fixed: Fresh greens, clean whites, natural meat tones, no golden overlay
Prompt used: Food variant from prompts/neutral_color_balance.md

---

3️⃣ 🎬 Cinematic / Fantasy Art

Problem: Heavy orange/teal grading — orange skin tones, teal shadows, sepia haze

Before (Orange/Teal)	After (Neutral Fantasy)

Preserved: Composition, characters, mood, detail
Fixed: Natural skin, neutral shadows, readable text, no orange cast
Prompt used: Cinematic variant from prompts/neutral_color_balance.md

---

🚀 Quickstart

📦 Installation by Tool

🐙 opencode git clone https://github.com/BrunosGits/sepia-be-gone.git ~/.config/opencode/skill/sepia-be-gone # Use: /sepia-be-gone 🤖 Claude Code git clone https://github.com/BrunosGits/sepia-be-gone.git ~/.claude/skills/sepia-be-gone # Use: /sepia-be-gone ⚡ Codex / Generic Agents git clone https://github.com/BrunosGits/sepia-be-gone.git .github/skills/sepia-be-gone # Use: /sepia-be-gone

🎯 Cursor # Copy CURSOR.md to: ~/.cursor/rules/sepia-be-gone.mdc # Or project: .cursor/rules/ # Use: @sepia-be-gone in chat 🌊 Windsurf # Copy WINDSURF.md to: ~/.codeium/windsurf/skills/sepia-be-gone/ # Use: /sepia-be-gone in Cascade 💻 VS Code Copilot # Add VSCODE.md content to: .github/copilot-instructions.md # Use: "Use sepia-be-gone skill" in chat

🎨 Any Image Generator

Works with DALL-E 3, Midjourney, GPT-4o, Gemini, Firefly, Stable Diffusion, Flux, etc.

cat prompts/neutral_color_balance.md | pbcopy
# Paste at end of your prompt

---

📋 The Prompt (Copy-Paste Ready)

📄 Full Prompt

→ prompts/neutral_color_balance.md

⚡ Short Add-on (append to any prompt)

Color grading: neutral daylight-balanced white balance, accurate colors, clean whites, neutral blacks, no yellow tint, no sepia, no orange cast, no warm filter, no vintage grading, no golden-hour lighting.

🚫 Negative Prompt (if supported)

yellow tint, orange cast, warm filter, sepia, amber lighting, vintage color grading, old painting, golden hour, aged paper, brown overlay, muddy colors, oversaturated orange, excessive warmth, cream whites, brown blacks

Midjourney: Add --no yellow tint, orange cast, warm filter, sepia, amber lighting, vintage color grading, old painting, golden hour, aged paper, brown overlay, muddy colors, oversaturated orange, excessive warmth, cream whites, brown blacks
Stable Diffusion: Paste into negative prompt box

🎯 When to Use

Click to expand trigger phrases

• "Remove the yellow tint" / "Remove yellow filter"
• "Make this less warm" / "Too warm" / "Fix warm colors"
• "Neutralize colors" / "Neutral color balance"
• "Clean whites" / "Whites look cream" / "Whites are yellow"
• "Remove orange cast" / "Too orange"
• "Remove sepia" / "Looks vintage" / "Old photo look"
• "Fix AI yellow filter" / "AI yellow tint"
• "5600K" / "Daylight balanced" / "Neutral lighting"

---

⚠️ Limitations & Workarounds

Limitation	Severity	Workaround
Generative editors distort text	🔴 High	For posters with exact prices/addresses: generate background only → add text in Figma/Canva/Photoshop
Deliberate warm/vintage style	🟡 Medium	Don't use — this skill forces neutral
Extreme orange casts (sunset)	🟡 Medium	May need 2-pass editing
Local color casts only	🟢 Low	Use generative inpainting for specific regions

---

📁 Repository Structure

sepia-be-gone/
├── 📄 README.md                          # This file (with inline examples)
├── ⚙️  SKILL.md                           # opencode skill (primary)
├── 🤖 AGENTS.md                          # Codex/opencode adapter
├── 🤖 CLAUDE.md                          # Claude Code adapter
├── 🎯 CURSOR.md                          # Cursor IDE rule (.mdc)
├── 🌊 WINDSURF.md                        # Windsurf Cascade skill
├── 💻 VSCODE.md                          # VS Code Copilot instructions
├── ⚖️  LICENSE                            # MIT
├── 📁 prompts/
│   └── 🎨 neutral_color_balance.md       # The reusable prompt
└── 📁 examples/
    └── 📁 images/
        ├── 📁 before/                     # Your yellow-tinted inputs
        │   ├── poster-yellow.jpg
        │   ├── food-warm.jpg
        │   └── cinematic-orange.jpg
        └── 📁 after/                      # Your corrected outputs
            ├── poster-neutral.jpg
            ├── food-neutral.jpg
            └── cinematic-neutral.jpg

---

🛠 How It Works

🎭 Generative Prompt Strategy

The prompt uses three mechanisms:

1. Explicit preservation directives — "Preserve: layout, text, logos, prices, addresses"
2. Positive color specification — "5600K daylight-balanced", "clean whites", "neutral blacks"
3. Negative constraint list — 15+ forbidden terms (yellow tint, sepia, golden hour, vintage, etc.)

This exploits how diffusion models attend to both positive and negative token guidance.

🎯 Target Color Temperature

Lighting Type	Kelvin	Use Case
Tungsten (warm)	3200K	❌ Avoid
Golden Hour	3500-4500K	❌ Avoid
Daylight Neutral	5500-5600K	✅ Target
Overcast Daylight	6500K	✅ Acceptable
Cool Studio	7000K+	⚠️ May overcorrect

---

🤝 Contributing

Found a better prompt variant? Discovered a new tool integration?
PRs welcome! Please read the skill files before modifying.

---

📜 License

MIT License — Free for personal and commercial use.
See LICENSE for full text.

---

🌟 Star this repo if it saved your images from the yellow filter!

Made with ☕ by BrunosGits