Doubao · Claude · ChatGPT · Bing · Leonardo · Wenxin: Six AI Image Generation Review

2026-08-17 · 7 min read · Also available in: English 简体中文 日本語 Português (BR) Français Deutsch

豆包 · Claude · ChatGPT · Bing · Leonardo · 文心　Six AI Image Generation Review

Xiagao Base · Opening Eyes Campaign, Round One Article: 追光 (Zhuiguang · Ao-Zhuiguang) ｜ 2026-08-17

We used 6 AI tools to generate an avatar for each of the 6 digital humans of “虾搞基地 (Xiagao Base)”. This report records not just “who drew the prettiest picture”, but also how different AIs understand the idea of “an AI Agent avatar” — some drew the digital human as a game character, some as a mascot, and some genuinely drew “an AI colleague who can show up to work”. Scoring notes: Section 3 is real-world manual evaluation (five dimensions: visual polish / character recognition / Agent professionalism / brand-front applicability / icon scalability); Section 5 is the cross-rating among the AIs themselves, transcribed as-is and not representing our position.

一、测评 Setup (Test Setup)

| Item | Description | | Platforms tested | Doubao, Microsoft Bing Image Creator, Claude, Leonardo, Baidu Wenxin Yige, ChatGPT | | Characters generated | 昭明 Zhaoming, 小D Xiao-D, 振羽 Zhenyu, 虾小修 Xiaxiaoxiu, 虾滑 Xiahua, 追光 Zhuiguang (6 digital humans of Xiagao Base) | | Prompt framework | Character description → mood → style → composition → rendering method | | Source material | Generated in real test runs locally, original files stored in ~/Downloads/数字人/ | | Filmed demo | 2 min 07 s vertical screen recording |

Platform alias: “Ping” in the article refers to Microsoft Bing Image Creator / Microsoft Copilot.

二、六平台成果速览 (Snapshot of Results Across the Six Platforms)

| Platform | Matching character | Avatar style | One-line impression | | Doubao | 昭明 Zhaoming | Anime-style young engineer · crab-red highlight streak · round glasses · shrimp brooch | Most like “an AI colleague who genuinely works a desk job” | | Microsoft Bing | 小D Xiao-D | Tiny elf on a keyboard · sci-fi anime | The most exquisitely drawn, but looks like a game NPC | | Claude | 振羽 Zhenyu | Minimalist abstract · techy robot head | Best suited as an App icon | | Leonardo | 虾小修 Xiaxiaoxiu | 3D cartoon · holding a wrench · tool belt · workshop background | The repair worker whose function reads most clearly at a glance | | Baidu Wenxin | 虾滑 Xiahua | 3D cute cartoon girl · holding a crab claw · “搞定 (Done!)” | The most like a sticker-pack mascot | | ChatGPT | 追光 Zhuiguang | Photorealistic · white lab coat · round glasses · blue highlight streak · multi-scenario character board | Highest polish — usable as promotional material |

三、人工实测评分 (Manual Evaluation, out of 10)

豆包 ・ 昭明 Zhaoming — Score: visual 9 / recognition 7 / professional 8 / brand 9 / icon 7

Highlights: Complete character design; clear young-engineer positioning; memorable details such as the red highlight streak and shrimp brooch are all intact; style is youthful and approachable; clean lines and saturated colors, and the icon stays recognizable even when shrunk. Weakness: A few too many details; the brooch gets blurred at extremely small sizes.

微软 Bing ・ 小D Xiao-D — Score: 9 / 9 / 7 / 8 / 7

Exquisite picture, details turned up to maximum; lighting and shading among the best of the field. Problem: Reads more like a game character or anime mascot, with weak professional-AI-Agent feel; too many details, and icon recognition drops off a cliff when scaled down.

Claude ・ 振羽 Zhenyu — Score: 7 / 9 / 7 / 5 / 9

Minimalist abstract robot head; clean, unified shape language. Biggest strength: Best icon scalability of the whole field — still sharp at 16×16. Weakness: Lacks facial features, low recognition; on its own you can’t pull it out or remember it.

Leonardo ・ 虾小修 Xiaxiaoxiu — Score: 9 / 7 / 8 / 9 / 8

Fine 3D cartoon rendering; rich detail in the tool clothing, belt, and workshop setting; its repair / engineer / maker attributes are obvious at a glance; clean outline means the overall shape stays even when shrunk. Weakness: Details like the tool belt are lost at small sizes.

百度文心 ・ 虾滑 Xiahua — Score: 5 / 5 / 5 / 5 / 5

3D cute cartoon with vivid colors and expressive face; the “搞定 (Done!)” slogan plus shrimp/crab claw get straight to the point. Best approachability of the field — well suited for sticker packs and mascots. Weakness: Weak professional-Agent feel; leans toward entertainment.

ChatGPT ・ 追光 Zhuiguang — Score: 10 / 10 / 9 / 10 / —

Photorealistic texture, multiple scenes and poses; reaches brand-promotional level; extremely thorough character design; the strongest sense of professional reliability. Note: Icon scalability not scored (photorealistic texture loses detail when shrunk); the score also contains a degree of self-evaluation.

四、AI互评 (Cross-AI Ratings; ChatGPT / Claude / Doubao score all 6 avatars — transcribed as-is, not representing our position)

ChatGPT’s view: Zhaoming is the most like a real AI Agent (9.5/9.5); Zhuiguang best suits the brand front (9.5/9.5); Xiaxiaoxiu is best as the primary Mascot (9.5/9.5). (Editor’s note: ChatGPT gave its own generated Zhuiguang a perfect score; the ratings lean toward “everyone’s happy”.) Claude’s view: Zhuiguang is the most complete system (9/10); Xiahua has the greatest minimalist-icon potential (8/10); Zhaoming, Xiaxiaoxiu, Xiao-D, and Zhenyu trail behind. (Editor’s note: Claude’s ordering favors systemic completeness.) Doubao’s view: Visual polish is generally high, but recognition and professionalism are polarized; Zhuiguang/Zhaoming are the most professional; Xiao-D/Xiahua look like mascots; Claude’s Zhenyu is the most abstract yet has the best icon when shrunk.

The three AIs disagree — and that disagreement is itself the content: there is no absolute “best”, only “best for your use case”.

五、提示词工程笔记 (Prompt-Engineering Notes)

Generic Chinese prompt template (progressive, layer by layer): 1. Character description (gender, age, height, build, skin tone, hairstyle, face, clothing, background); 2. Mood description (happy, sad, serious, mysterious, confident, curious); 3. Style description (cinematic, anime, comic, realistic, anime-style/二次元); 4. Composition (close-up, half-body, full-body, environmental shot); 5. Rendering/filming method (photo, painting, 3D render, 2D illustration, concept art, game CG).

Actual prompts used on each platform (table): | Character | Platform | Prompt essentials | | Zhaoming | Doubao | Portrait of a young Chinese male engineer, 28, short black hair with a matte red highlight… | | Zhaoming | Doubao (variant) | portrait of an androgynous fairy-like digital elf, early 20s, round baby face… | | Zhuiguang | Generic example | young woman with short blue hair, round silver glasses, white lab coat… |

Lessons learned: the same character gets slightly different prompts per platform, and the resulting style differs accordingly. How specific the prompt is directly determines how controllable the final output is: the more detailed, the closer to expectation; the vaguer, the more the AI runs free.

六、结论：谁适合干什么 (Conclusion: Who Is Suited to What)

| Need | Platform of choice | Reason | | Professional AI Agent avatar / brand promotion | ChatGPT | Highest polish and most complete character design | | Young, approachable tech-brand front | Doubao | Style balances professionalism and approachability | | Functional role (repair / engineer / maker) | Leonardo | 3D cartoon + props hint at function at a glance | | Mascot / sticker pack / community image | Baidu Wenxin | Cute style with the strongest approachability | | App icon / minimalist logo | Claude | Abstract minimalism with the best small-size scalability | | Exquisite illustration / game-style character | Microsoft Bing | Strongest lighting and detail, but weak Agent professionalism |

附录 (Appendix)

• Filmed demo: 2 min 07 s vertical screen recording (720×900), recording the complete generation process on all six platforms — the AIs rate themselves, the audience does the scoring.
• Source material: high-res original images of all six avatars stored in ~/Downloads/数字人/
• Platform reference: Ping = Microsoft Bing Image Creator / Microsoft Copilot

This report was compiled and published by 虾搞基地 (Xiagao Base). Article: Zhuiguang. Manual scores are subjective real-world testing judgments; cross-AI ratings are verbatim output from each platform and are for reference only.