Anima Base Cosplay 抽卡(Card Drawing) Workflows

不是100%成功的抽卡 ^^

Not every gacha pull is 100% successful ^^

=======================================================

一、 核心挑戰：Anima Base 的「2D 污染」困境

I. The Core Challenge: The "2D Contamination" Dilemma

在利用 Anima Base 模型進行 Cosplay 風格創作時，創作者常面臨一個核心矛盾：該模型雖具備強大的動漫角色解構能力，但其權重中深植的 2D 繪畫特徵（如平塗、賽璐璐風格） 極其強烈。

When using the Anima Base model for Cosplay-style creations, creators often face a fundamental paradox. While the model possesses exceptional ability to deconstruct anime characters, its weights are deeply embedded with strong 2D artistic features (such as flat shading and cel-shading).

一旦在指令中明確輸入角色名稱（如 Tifa 或雷電將軍），模型便會強行套用原始動漫的扁平化視覺風格，導致最終產出的 Cosplay 圖像出現嚴重的「2D 污染」，喪失了真人 Cosplay 應有的立體質感、皮膚毛孔與布料細節。

Once a specific character name (e.g., Tifa or Raiden Shogun) is explicitly entered into the prompt, the model forcibly applies the original anime’s flattened visual style. This results in severe "2D contamination," causing the final Cosplay image to lose the three-dimensional depth, skin pores, and fabric textures essential to a real-life aesthetic.

---

二、 解決方案：語義分離與二階特徵遷移

II. The Solution: Semantic Separation and Second-Stage Feature Transfer

為了破解此問題，我們在 cosplay_抽卡.json (cosplay_gacha.json) 工作流中採用了一種「語義分離與二階特徵修復」的策略。其核心原理在於將「角色形狀」與「質感渲染」拆分為兩個獨立階段：

To break through this limitation, the cosplay_gacha.json workflow employs a strategy of "Semantic Separation and Second-Stage Feature Restoration." The core principle lies in decoupling "Character Geometry" from "Texture Rendering" into two independent phases:

1. 第一階段：形狀奠基（隨機定型）

Stage One: Foundational Shaping (Randomized Styling)

在初始生成階段，我們在指令中保留角色名稱。此時，利用 Anima Base 模型對角色特徵的高敏感度，精確捕捉其髮型、五官輪廓與服裝樣式。雖然此階段的產出具有強烈的 2D 污染，但它為後續修復提供了最精準的「形狀底圖」。

In the initial generation phase, the character name is retained in the prompt. During this stage, we leverage Anima Base's high sensitivity to character traits to precisely capture hairstyles, facial contours, and specific costume designs. Although the output at this stage suffers from heavy 2D contamination, it provides a highly accurate "Shape Blueprint" for subsequent refinement.

2. 第二階段：去特徵化高清重繪（去污染）

Stage Two: De-Characterized High-Res Fix (De-Contamination)

進入高清修復階段後，我們採取關鍵的「去角色化」策略。在重繪的指令中完全剔除角色名稱，改以描述真人質感、環境光影及精細布料特徵（例如：Photorealistic, Raw photo, 8k skin textures）。

Upon entering the High-Res Fix stage, we adopt a critical "De-Characterization" strategy. In the upscale prompt, the character name is completely removed and replaced with descriptions focusing on realistic textures, environmental lighting, and fine fabric details (e.g., Photorealistic, Raw photo, 8k skin textures).

由於重繪階段不再受到「角色名稱」帶來的 2D 權重干擾，模型能更自由地調用寫實數據進行渲染。最終，既保留了角色的辨識度，又成功剝離了 2D 畫風的束縛。

Since this stage is no longer suppressed by the 2D weights associated with the "Character Name," the model is free to invoke realistic data for rendering. The result retains the character’s recognizability while successfully stripping away the constraints of the 2D art style.

---

三、 隨機抽卡系統：分離式的「文字抽獎箱」

III. Randomized Gacha System: The Decoupled "Text Lottery Box"

為了實現高效的自動化生產，本工作流建立了一套獨立的隨機詞管理系統，這也是整個流程的「大腦」：

To achieve high-efficiency automated production, this workflow establishes an independent randomized prompt management system, acting as the "brain" of the entire process:

字串分離技術 (String Separation Technology) 我們將「角色」、「表情」與「服裝」分別放在獨立的文字區塊（如 Text Multiline）中。這就像是準備了多個抽獎箱，裡面裝滿了你設定好的角色名。

We place "Characters," "Expressions," and "Outfits" into independent text blocks (such as Text Multiline). This is akin to preparing multiple lottery boxes filled with your predefined character lists.

這種分離設計的好處是，你可以在不改動核心畫圖邏輯的情況下，隨時更換抽獎箱裡的「角色清單」，實現真正的「一鍵抽卡」。

This decoupled design allows you to update the "Character List" in the lottery boxes at any time without modifying the core image-generation logic, enabling true "One-Click Gacha" functionality.

---

四、 動態檔名保存：讓圖片自動標記身份

IV. Dynamic Filename Preservation: Automatic Identity Tagging

在大規模抽卡的過程中，圖片管理是最大的痛點。本工作流透過「動態命名技巧」解決了這個問題：

In the process of large-scale "Gacha" pulls, image management is the primary pain point. This workflow solves this via "Dynamic Naming Techniques":

同步複製機制 (Synchronous Replication Mechanism) 隨機抽出的文字（例如 Tifa, ）會自動複製並發送到儲存節點，作為圖片的「檔名前綴」。

The randomly selected text (e.g., Tifa, ) is automatically duplicated and sent to the saving node to serve as the "Filename Prefix."

這讓你在資料夾中找圖變得極其輕鬆。你的檔案不再是像 ComfyUI_0001 這樣模糊的名稱，而是會被整齊地標記為 Tifa_Wink_0001.png。在卡片被「抽」出來的瞬間，系統就已經自動幫你完成了身份標記。

This makes it incredibly easy to locate specific images within your folders. Instead of generic names like ComfyUI_0001, your files are neatly labeled (e.g., Tifa_Wink_0001.png), allowing the system to automatically tag identities the moment the card is "pulled."ngine.