About This Article
This article is organized using generative AI, referencing OpenAI's public documentation and actual papanda925 images. The papanda925 Character Sheet v1 and fixed rules use GitHub as the source of truth.Verification Status: 🧪 Character Sheet baseline and fixed rules verified / 3-generation comparison to be conducted next time
You managed to generate one character you really like using AI.
A few days later, when you ask it to "generate the same character in a different pose,"it looks similar, but slightly like a different person.
In image generation, this is surprisingly troublesome. At papanda925.com, I want to use the same panda across blogs, manga, and social media, so I want to avoid changes to the face and body type every time.
The workflow I tried this time is simple.
Your favorite image → Create a Character Sheet → Set it as the official reference → Change only expressions, poses, and scenes
The mindset is:
Change the acting, not the actor.
Change the acting, not the actor.
.
- 1. Start with your "favorite image"
- 2. Review the generated Character Sheet
- 3. Separate "what to fix" from "what can change"
- 4. Generating different poses from the Character Sheet
- 5. Use "Editing" rather than "New Generation" to further reduce drift
- 6. Using GitHub as the official repository
- Summary: Start from one image and cultivate the baseline
- References
1. Start with your "favorite image"
You don't need to prepare multiple front, side, and expression views from the beginning.
If you have even one image that makes you think, "I want to keep using this character," you first provide it as a reference image.
Then, you use the next prompt to create a Character Sheet.
添付画像のキャラクターを、今後使用する正式なキャラクターとして整理したいです。 このキャラクターを「papanda925」とします。 添付画像を最優先の参照画像として、同じキャラクターのCharacter Sheetを作成してください。 1枚のシートに次をまとめてください。 ・正面 ・斜め ・横向き ・考えている表情 ・驚いている表情 ・うれしい表情 ・基本的な立ち姿 ・PCを操作している姿 ・基本カラーパレット 新しいキャラクターをデザインしないでください。 元画像の顔、耳、目、アイパッチ、鼻、口、頬、頭身、体型、基本配色、イラストタッチを維持してください。 これは新キャラクターのデザインではなく、添付された同一キャラクターの設定資料です。
2. Review the generated Character Sheet
Here is the official reference created this time.
It combines the front view, three-quarter view, profile, expressions, PC work, and color palette into a single image.
However,when there is only one source image, the AI guesses parts that are not visible, such as the profile.
Therefore, the Character Sheet is not finalized on the first try, and any unnatural parts are corrected.
For example, if the profile looks too thin, you fix it like this.
Character Sheetを修正してください。 元のpapanda925画像を最優先してください。 横向きの顔が細くなっているので、元画像と同じ丸い顔の印象に戻してください。 耳も少し大きいので、元画像と同じ大きさにしてください。 それ以外は変更しないでください。
The version you are satisfied with here is established as Character Sheet v1.
3. Separate "what to fix" from "what can change"
At papanda925, fixed rules are maintained in CHARACTER_SOURCE.md on GitHub.
| Fixed Elements | papanda925 Criteria |
|---|---|
| Face and belly | Cream tones #FFF8F0 |
| Ears, eye area, limbs | Dark brown tones #4A3425 |
| Eyes and nose | #222222 |
| Cheeks | #FFD7C2 |
| Accents | #E6F3FE |
| Body type | Large head + small body, 2-to-3 heads tall proportions |
| Art style | Gentle, flat illustration style |
On the other hand, backgrounds, props, expressions, poses, and article themes are allowed to change.
In other words,the approach is to fix the elements related to the character's ID while changing only the performance for that specific instance.
4. Generating different poses from the Character Sheet
Next, you pass the Character Sheet as a reference image and specify only the scene.
papanda925 thinking
添付した papanda925 Character Sheet v1 を正式なキャラクター参照として使用してください。 新しいパンダをデザインしないでください。 【SCENE】 papanda925がノートPCの前に座り、片手をあごに当てて考えている。 「原因はどこだろう？」と考えている雰囲気。 【CHANGE ONLY】 変更するのは、表情、視線、腕の位置、ポーズ、小物、背景だけです。 顔、耳、目、アイパッチ、鼻、口、頬、頭身、体型、基本配色、イラストタッチは維持してください。 Change the acting, not the actor.
papanda925 surprised
The Character Sheet remains the same, and only the SCENE changes.
【SCENE】 同じpapanda925がPCを操作している。 画面にエラーが表示され、驚いてモニターを見ている。 口を開け、両手を少し上げている。 【CHANGE ONLY】 表情とポーズだけを変更する。 キャラクターそのものは変更しない。
papanda925 explaining
【SCENE】 同じpapanda925がホワイトボードの前で、指し棒を使って解決方法を説明している。 理解できてうれしそうな表情。 【CHANGE ONLY】 変更するのは、表情、腕、指し棒、ホワイトボードだけ。 papanda925本体のデザインは変更しない。
With this approach, the core instruction shifts from "draw a panda" to"what papanda925 should be made to do this time."
5. Use "Editing" rather than "New Generation" to further reduce drift
Once you have an image you like, instead of recreating it from scratch every time, you can also use a method of editing that image.
この画像を編集してください。 papanda925本体、PC、机、背景は維持してください。 右手だけを上げて、PC画面を指すポーズに変更してください。 それ以外は変更しないでください。
OpenAI's image generation guide also introduces a method of specifically designating what to change and what to keep, accumulating small corrections.
Rather than "recreating everything," keepingChange onlyin mind is more suitable for character management.
6. Using GitHub as the official repository
If you keep the Character Sheet only inside a chat, you will later wonder, "Which one was the latest version?"
Therefore, at papanda925, the official reference was placed on GitHub.
papanda925 character ├─ papanda925-character-sheet-v1.png ← 画像の正本 └─ CHARACTER_SOURCE.md ← 色・形・固定ルール
Instead of relying on the AI's memory every time an image is generated,this creates a state where the exact same reference can be accessed every time.
Summary: Start from one image and cultivate the baseline
The workflow consists of the following 5 stages:
Choose a single character image you like
Create a Character Sheet from that single image
Correct parts distorted by guesswork
Save the finalized Character Sheet as the official reference
Generate images thereafter with "fixed character + modified acting only"
Rather than writing long prompts every time,it is more important for the human side to maintain the standard that "this character is correct."
Whether tinkering with PowerShell, explaining Excel, or being surprised by an error, the same papanda925 is right there.
Change the acting, not the actor.
Change the acting, not the actor.
Next time, I will generate papanda925 three times using the same Character Sheet and conditions,and compare them side by side to see how much character drift can actually be suppressed.
