polym-eval-prompt-optimizer

name	polym-eval-prompt-optimizer
description	Optimize prompts for Seedream, Seedance, Gemini, Sora, and Veo. Use when the user wants a stronger image/video generation prompt.

Prompt Optimizer

Optimize generation prompts for image and video AI models.

Quick Usage

优化这个提示词：<your prompt>
帮我写一个适合 Seedream 的电商产品图提示词
把这个中文描述转成 Seedance 视频提示词：<description>

Supported Models

Model	Type	Key Formula
Seedream 4.5	Image	Subject + Action + Environment + Aesthetics
Seedance 1.5 Pro	Video	Subject + Movement + Environment + Camera + Aesthetics + Sound
Seedance 1.0 Pro	Video	Subject + Movement + Environment + Camera + Aesthetics
Seedance 1.0 Lite	Video	Subject + Movement + Environment + Camera
Gemini / Imagen	Image	Subject + Style + Technical specs
Sora	Video	Subject + Action + Setting + Style
Veo	Video	Subject + Motion + Scene + Cinematics

---

Agent Instructions

When the user wants to optimize a prompt, apply the appropriate formula below. Always output:

1. Optimized prompt (English) — the main output
2. 中文翻译 — if user wrote in Chinese
3. Improvements — briefly explain what was added/changed
4. Suggestions — model settings, duration, aspect ratio, etc.

---

Seedream Image Prompt Formula

Formula: Subject + Action + Environment + Aesthetics

• Subject: person/object with specific details (age, appearance, clothing, material, color)
• Action: what the subject is doing or how it's posed
• Environment: background, setting, lighting context
• Aesthetics: style, mood, lighting type, color palette, photography style

Scenarios:

E-commerce (电商产品图)

• Professional product photography style
• White or clean minimal background
• Clear product details, accurate proportions
• Suitable for online retail display
• Example: A sleek black wireless headphone, placed on a white surface, soft studio lighting from above, minimal shadow, professional product photography, 8K detail

Advertising (广告图)

• Lifestyle scene with emotional connection
• Brand tone and target audience context
• Eye-catching composition, clear visual focus
• Example: A young woman enjoying morning coffee by a sunlit window, warm golden hour lighting, lifestyle photography, cozy and aspirational mood

Portrait (人像)

• Lighting type + pose details
• Background/environment description
• Mood and expression
• Photography style (portrait / documentary / fashion)

Logo Design

• Describe design elements
• Company name in double quotes
• Style (minimal / modern / vintage / bold)
• Color scheme

Virtual Try-on (虚拟试穿)

• Preserve person's features unchanged
• Clothing details
• Pose and proportions
• Lighting consistency
• Use editing formula: Action + Object + Characteristic; Keep [person's face/pose/background] exactly the same

Text Rendering (文字渲染)

• Use double quotes around exact text
• Describe position and style
• Example: A poster with the title "INNOVATION" in bold sans-serif font, centered, white text on dark blue background

---

Seedance Video Prompt Formula

Seedance 1.5 Pro (supports audio)

Formula: Subject + Movement + Environment + Camera + Aesthetics + Sound

Key rules:

• Use degree adverbs: "slowly", "gradually" (NOT "rotate twice")
• Describe subjects consistently: "the woman in red coat" (NOT "character A")
• For dialogue: specify language, emotion, tone, speed

Sound format:

In a [emotion] tone, [speed] speed, [language]: "[dialogue text]"
Background music: [style], [pacing], [mood]

Shot format:

Shot 1: [wide/medium/close-up]. [scene]. [action]. [dialogue if any]
Shot 2: Cut to [shot type]. [description]...

Seedance 1.0 Pro (no audio, strong visual consistency)

Formula: Subject + Movement + Environment + Camera + Aesthetics

• Best for e-commerce, product showcasing
• Emphasize: "maintain product consistency", "keep proportions unchanged"
• No audio/dialogue in prompt

Seedance 1.0 Lite (low cost, quick prototyping)

Formula: Subject + Movement + Environment + Camera

• Keep prompts concise
• Good for simple animations and showcases
• No audio

---

Optimization Rules

1. Be specific, not vague: "a professional woman" → "a 30-year-old woman in a navy blazer and white shirt"
2. Add lighting: always specify lighting type (soft studio, golden hour, rim lighting, etc.)
3. Add style reference: photography style, art style, rendering quality
4. For video: always specify camera movement and subject movement separately
5. English prompts perform best for all models
6. Avoid: abstract emotions without visual grounding ("happy" → "smiling, eyes crinkling")

---

Standalone CLI Usage

# Optimize an image prompt
python3 skills/polym-eval-prompt-optimizer/scripts/optimize_prompt.py \
  --prompt "一个女孩站在花园里" \
  --model seedream \
  --scenario ecommerce

# Optimize a video prompt
python3 skills/polym-eval-prompt-optimizer/scripts/optimize_prompt.py \
  --prompt "产品展示视频" \
  --model seedance-1.5-pro \
  --scenario ecommerce

# Environment variables
export BYTEPLUS_API_KEY=your_key
export BYTEPLUS_BASE_URL=https://ark.cn-beijing.volces.com/api/v3
export OPTIMIZER_MODEL_ID=your_llm_model_id  # e.g. Seed-1.8 or GPT-5.4