Higgsfield 知識庫

圖像提示思考邏輯|由想法到優化提示

步驟 1:意圖 → 模型(官方 Picking flow 原文)skills-catalog

### Image — auto-pick

1. **Brand product visual (Pinterest pin, lifestyle, hero banner, ad pack, virtual try-on, restyle)** → use `higgsfield-product-photoshoot` instead. NOT this skill.
2. **Branded ad image with presenter avatar + product (Marketing Studio shape with RAG over user assets)** → Marketing Studio Image.
3. **Soul Character (reference id from `higgsfield-soul-id`)** → Soul 2.0 for stills; Soul Cinema (`soul_cinematic`) for cinematic vibe.
4. **New original person — UGC creator, influencer, editorial, fashion** → Soul 2.0.
5. **Cinematic still frame** → Soul Cinema (`soul_cinematic`).
6. **Character sheet, one-shot face from reference photos, or face edit on a real photo** → Seedream 5.0 Pro (`seedream_v5_pro`).
7. **Locations / environments / no-people scenes** → Soul Location.
8. **Logo, icon, vector-like illustration, brand mark, controlled-palette graphic** → Recraft V4.1. Use `--model_type vector` for vector-style output.
9. **Cartoon or illustrated characters, heavily textured photos** → Nano Banana 2 (`nano_banana_flash`).
10. **Default for everything else** → GPT Image 2.5. General generation, editing, design, UI, banners, product concepts, anything with on-image text.

### Image — only when asked

- **Cheaper or faster:** Nano Banana 2 Lite (`nano_banana_2_lite`) for quick reference edits; Z Image for fast drafts.
- **Named by the user:** Nano Banana Pro, Soul Cast, Soul Cinema Studio, Seedream 5.0 Lite, Seedream 4.5, Flux 2, Flux Kontext Max, Kling O1 Image, Grok Image, GPT Image 2, Cinema Studio Image 2.5, Auto, and the unprofiled models listed above.
flowchart TD
  A[圖像需求] --> B{品牌產品圖?<br>Pinterest/生活照/hero/試穿}
  B -->|是| PP[higgsfield-product-photoshoot<br>→ GPT Image 2 模板]
  B -->|否| C{品牌廣告 + 代言人 + 產品?}
  C -->|是| MS[Marketing Studio Image]
  C -->|否| D{已有 Soul 角色?}
  D -->|是| S2[Soul 2.0 / Soul Cinema]
  D -->|否| E{新原創人物<br>UGC/時裝/editorial?}
  E -->|是| S2b[Soul 2.0]
  E -->|否| F{電影感靜幀?}
  F -->|是| SC[Soul Cinema]
  F -->|否| G{角色表/參考照臉部編輯?}
  G -->|是| SD[Seedream 5.0 Pro]
  G -->|否| H{無人場景/地點?}
  H -->|是| SL[Soul Location]
  H -->|否| I{Logo/圖示/向量?}
  I -->|是| RC[Recraft V4.1]
  I -->|否| J{卡通/插畫角色?}
  J -->|是| NB2[Nano Banana 2]
  J -->|否| GPT[GPT Image 2.5(預設)]

步驟 2:選提示結構

情況 結構 模型頁
一般生成 主體 + 場景 + 風格 + 鏡頭 + 光線 + 媒材(≤ ~200 tokens) GPT Image 2.5
圖上有文字/UI/海報 把要出現的文字原樣放在引號內,指明位置與字體風格 GPT Image 2.5、GPT Image 2
編輯現有圖 只寫「要改甚麼 + 要保留甚麼」 Nano Banana Pro、FLUX Kontext
人物/時裝/UGC 人物 + 服裝 + 場景 + 拍攝感(手機/雜誌) Soul 2.0
多參考合成 指明每張參考的角色(人物/產品/場景) Seedream 5.0 Pro、FLUX 3 Image
快速起稿 簡短提示,在 Z Image 迭代,方向確定後換預設模型 Z Image

來源:skills-prompt-eng skills-cookbook;每個模型的官方公式與逐字範例見模型頁。

步驟 3:參數

步驟 4:迭代

  1. 先產 1 張確認方向,再擴量。blog:claude-higgsfield-mcp-creative-studio
  2. 每次只改一個變數。blog:how-to-avoid-distortions-ai-videos
  3. 被拒(nsfw/ip_detected)→ 移除真實名人、商標角色、色情字眼。skills-prompt-eng

本頁來源(存取日期 2026-10-08)

ID 標題 發佈日期 URL
skills-catalog skills — higgsfield-generate/references/model-catalog.md 2026-09-26 https://github.com/higgsfield-ai/skills/blob/main/higgsfield-generate/references/model-catalog.md
skills-prompt-eng skills — higgsfield-generate/references/prompt-engineering.md 2026-09-26 https://github.com/higgsfield-ai/skills/blob/main/higgsfield-generate/references/prompt-engineering.md
skills-cookbook skills — COOKBOOK.md 2026-09-26 https://github.com/higgsfield-ai/skills/blob/main/COOKBOOK.md
cli-models higgsfield-ai/cli — MODELS.md(官方 CLI 模型參數表,commit dc7e2d2, 2026-09-11) 2026-09-11 https://github.com/higgsfield-ai/cli/blob/main/MODELS.md
blog:claude-higgsfield-mcp-creative-studio claude higgsfield mcp creative studio 2026-09-02 https://higgsfield.ai/blog/claude-higgsfield-mcp-creative-studio
blog:how-to-avoid-distortions-ai-videos how to avoid distortions ai videos 2026-07-18 https://higgsfield.ai/blog/how-to-avoid-distortions-ai-videos