InternS1 Spring Horse 2026 Retrospective: 9 AI Agents Collaborating to Create 20 4K Posters

4 min read
AIAgent TeamsCreativeInternLM

During the Spring Festival period, completed an interesting experimental project: using a team of 9 specialized AI Agents collaborating to create Year of Horse CNY greeting posters for 12 InternLM open-source project brands.

Project Goals

Core objectives of the Spring Horse 2026 project:

  • Create Year of Horse themed posters for 12 InternLM series brands (InternLM, XTuner, LMDeploy, MinerU, etc.)
  • Output 20 4K-level images + 5 videos
  • Quality score ≥ 4.0/5
  • Zero high-risk content

Division of 9 AI Agents

Multi-phase parallel team architecture:

Team Lead (Coordinator)
├── Phase 1 (Parallel)
│   ├── Director (Strategy) — Creative brief + final selection
│   ├── Writers (10 styles) — 40 image + 12 video concepts
│   └── Asset-Gatherer — 12 brand logos
├── Phase 3 (Parallel)
│   ├── Image-Gen — 20 4K posters
│   └── Video-Gen — 5 keyframes
└── Phase 4 (Parallel)
    ├── Quality-Reviewer
    ├── Virality-Reviewer
    └── Risk-Reviewer

Key Role Responsibilities

Director

  • Create master creative brief including theme positioning, visual style, copywriting direction
  • Define risk control rules (avoid political sensitivity, false promises, etc.)
  • Final selection and quality control

Writers (10 Copywriters)

  • Each writer responsible for 4 image concepts + 1-2 video concepts
  • Diverse styles: festive, tech, Chinese style, ink painting, etc.
  • Proactively identify and correct deviations from director's brief

Image-Gen (Image Generator)

  • Call Jimeng AI API to generate 5504×3072 / 4096×4096 high-resolution images
  • Handle brand logo overlay and layout optimization
  • Output 20 finished posters

Three-Track Review Team

  • Quality-Reviewer: Assess visual detail, composition, color, text legibility
  • Virality-Reviewer: Evaluate viral potential, emotional resonance, social attributes
  • Risk-Reviewer: Screen for political sensitivity, false promises, sexual innuendo, etc.

Final Results

MetricTargetActualAchievement
Image Count2020100%
Resolution4K5504×3072 / 4096×4096Exceeded (5K+ level)
Brand Coverage12 brands12/12 full coverage100%
Quality Score≥4.0/54.76/5Exceeded
High-Risk Content00Achieved

Incomplete Items

Originally planned to generate 5 videos, but due to Jimeng Seedance 2.0 API's anti-scraping system (ret=1019 error), only completed video keyframes and prompts. Video generation requires manual completion in browser.

Collaboration Highlights

  1. High Parallel Efficiency - Phase 1 three-track parallel, Phase 3 image/video parallel, Phase 4 three-track review parallel, significantly improving overall efficiency

  2. Writer→Director Closed Loop - Writer team proactively identified deviations from director's brief (e.g., missing Qwen brand, changing champagne to tea), director promptly released v1.1 revision

  3. Risk Forward - Director defined risk control rules in planning phase, strictly executed in generation phase, zero high-risk content in review phase

  4. Strict Quality Control - Three-track review independently scored and aggregated, final average 4.76/5, some posters achieving 5.0/5

Tech Stack

  • Agent Framework: Claude Code Agent Teams
  • Image Generation: Jimeng AI (Tusheng 2.0)
  • Video Generation: Jimeng Seedance 2.0 (keyframes only)
  • Logo Processing: Automated SVG/PNG overlay
  • Storage: Alibaba Cloud OSS

Lessons Learned

Success Factors

  1. Agent Specialization - Each agent with clear responsibilities, avoiding quality instability of generalist agents
  2. Parallel + Pipeline Combination - Creative phase parallel, generation phase pipeline, review phase parallel
  3. Risk Forward - Define rules in planning phase, avoid rework

Room for Improvement

  1. Version Sync Mechanism - Need more timely notification to downstream agents when director revises creative brief
  2. API Fault Tolerance - Need multiple backup APIs to avoid single point of failure (no backup when Jimeng blocked)
  3. Human Intervention Points - Key decision points (like final selection) still require human oversight

All 20 posters uploaded to InternS1 IP Creative Hub, covering:

  • InternLM (LLM)
  • XTuner (Efficient Fine-tuning)
  • LMDeploy (Inference Engine)
  • MinerU (Document Parsing)
  • Lagent (Agent Framework)
  • OpenCompass (Evaluation)
  • MindSearch (Deep Search)
  • SlideAgent (PPT Generation)
  • Qwen (Tongyi Qianwen)
  • Shusheng Puyu
  • InternLM Practical Camp
  • Shanghai AI Laboratory

Project Links

This project demonstrates the huge potential of AI Agent Teams in creative production. Through proper division of labor and collaboration mechanisms, complex creative projects can be efficiently completed.