Postmortem: Totally Accurate Succubus Simulator

Overview

Totally Accurate Succubus Simulator (TASS) is an erotic game utilizing AI-generated illustrations. The game contains 980 CGs, supports English, Japanese, and Chinese, and runs entirely in the browser. Developed solo, the project went from concept to release in about 10 days. This rapid turnaround was made possible by a development approach that minimized regression and leveraged AI to its fullest. In this article, I’ll reflect on the development process, key design decisions, and the technologies used.

Why Make This Game?

Previously, I had been posting erotic AI-generated art on Pixiv. While moderately successful—reaching about 7,000 followers and 3,000–4,000 views per post in three months—I found it difficult to differentiate myself by images alone. As technical barriers to high-quality illustration continue to fall, anyone can now generate good art with minimal effort. So, instead of focusing on AI art as a final product, I decided to use it as material for something else: a game.

Economically speaking, when the “price” (difficulty) of a commodity approaches zero, it makes sense to create complementary goods. While I might not stand out as an AI artist, perhaps I could as a game developer using AI art as my medium.

Project Overview

• Timeframe: 10 days, solo development
• Goal: Experiment with visual novel production methods using AI-generated art
• Specs: 980 AI illustrations, trilingual (EN/JA/ZH), browser/mobile support
• Results: 15,000 views on itch.io in the first 19 days after release

Tools Used

• Game engine: JavaScript + HTML (no frameworks)
• Prototyping: Python
• Image generation: A1111

Development Timeline

• Day 1: Generated concept images, prototyped rule system
• Day 2: Extracted rules into pure functions
• Day 3: Ported rules to JS, built mobile-friendly UI
• Day 4: Integrated UI and rules, added sample images
• Day 5: Developed semantic image retrieval system
• Day 6: Improved UI (image zoom, keyboard support)
• Day 7: Wrote story content
• Day 8: Implemented multilingual support
• Day 9: Generated images
• Day 10: Final upload

Technologies Introduced

• Dependency-based scheduling: Only began work on a task once prerequisites were complete
• Functional programming: All game rules implemented as pure functions
• Automatic prompt generation: Scripted parameter-based bulk prompt creation for AI art
• Semantic image retrieval: Mapped images to a vector space for dynamic, context-aware access

Design Decisions

• Minimal text, story told through images and rules
  • Reason 1: Not enough manpower to write extensive scenarios
  • Reason 2: A concise narrative with abundant images offers a better user experience
• No frameworks, hand-coded in vanilla JavaScript
  • Reason 1: Most VN tools target non-programmers and can slow down development for complex logic
  • Reason 2: Standard VN engines don’t align with my logic-centric, non-scripted structure
• No gallery mode
  • Reason 1: Players can download the game file and view all images directly if they want
  • Reason 2: Using WebP (instead of AVIF) made bulk image loading slower, so a gallery would add little benefit
• Browser/mobile support
  • Reason 1: Games requiring installation are a barrier to play
• Reason 2: While the PC adult game market is saturated, mobile-friendly erotic games remain rare

What Went Well

• No significant bugs or regressions occurred during development
• Functional programming enabled a focus on architecture and made AI-assisted implementation easier
• The AI image generation pipeline was highly automated, requiring only about 10 seconds of manual work per finished image
• The semantic retrieval system allowed for unlimited image additions without logic changes
• The codebase was modular and reusable for future projects

Areas for Improvement

• The protagonist’s appearance in generated images was inconsistent, leading to a high rate of unusable outputs; more specific prompt engineering could improve yield
• Despite omitting a gallery feature, player demand for it remained
• Not saving the random seed made it difficult to reproduce specific playthroughs

Next Steps

In future articles, I’ll dive deeper into each technology and stage of development. No worries if you’re not an expert—I’ll keep explanations clear and accessible.