Skip to content

Visual Preamble Exploration — Brief for a New Chat

Copy everything below the --- into a new Claude.ai chat.

I'm building Sync, a decision-making game where players face scenarios and make choices. Part of the experience is a reactive campaign mode where each chapter opens with a preamble — an atmospheric narrative scene that sets the mood before the player makes a decision. Right now preambles are text-only: a sequence of prose and dialogue blocks rendered one at a time with a typewriter effect, 1-second pause between blocks. It reads like the cold open of a TV episode.

Playtester feedback says the preambles work emotionally — they set the mood, build tension, introduce characters. But one player flagged: "I'm a slow reader, I felt pressure that people were waiting for me." Another said "I'd love to turn this into a comic strip." The text feels like a wall for some players.

The Goal

Explore whether we can generate visual accompaniment for preamble blocks — something like a graphic novel / comic strip experience. Not replacing the text, but complementing it. When a character appears, show their face. When a scene is set, show the room. When something dramatic happens, show the moment.

Constraints

  1. Cost target: under $0.10 per preamble for images. A preamble has 7-10 blocks. If every block needed an image, that's 1¢ per image. If only key moments need images, maybe 2-3 images per preamble at 3¢ each.

  2. Character consistency: A preamble might feature the same character (e.g., "Rebecca Thorne, CFO") multiple times across multiple chapters. She needs to look the same every time. This is the hardest part — naive image generation produces a different person each time.

  3. Scene consistency: The setting (e.g., "14th floor conference room at Meridian Health") should feel continuous across chapters.

  4. Speed: Preamble generation already takes ~15 seconds. Adding image generation can't make it painfully slow. Parallel generation is fine. Ideally images complete by the time the player finishes reading the first few blocks.

  5. Style: Graphic-novel or comic style feels right — illustrated, stylized, not photorealistic. Photorealistic faces hit uncanny valley and are more expensive.

Current Tech Stack Context

  • Anthropic Claude Sonnet generates the preamble text (JSON array of prose/dialogue blocks)
  • Each reactive campaign has a master doc that defines the world, player, and a recurring cast (4-6 NPCs with names, roles, personalities, and a "narrativeColor" like blue/amber/emerald)
  • Preambles reference cast members by name
  • Stored in Supabase Postgres; images would be stored as URLs pointing to... something (S3? Cloudflare R2?)

What I've Looked At

  • ComicMaker.ai and similar consumer products: too expensive for per-preamble generation at scale
  • Replicate.com: runs open-source models like Flux, SDXL — cheaper (~$0.01-0.03/image) but need to solve character consistency
  • Midjourney API: not officially available; quality is high but cost is unclear

Questions I Need You to Help Me Answer

  1. What's the right technical architecture? Probably Claude generates image prompts from the preamble text (captions, character references, scene descriptions), then something like Replicate generates the images. Walk me through the options.

  2. Character consistency — what are the real solutions?

  3. IP-Adapter + reference images? (generate one reference per cast member, reuse)
  4. LoRA training per cast member? (expensive, but more consistent)
  5. Seed locking with detailed description prompts?
  6. Character-consistent models like Flux Schnell + IP-Adapter?

  7. What actually works in 2026 at <$0.05/image?

  8. Cost math for realistic scale: If I have 100 users each playing a 5-chapter campaign with 3 images per preamble, that's 1,500 images. At what cost per image does this become viable? Is there caching I can do (same cast, same scene across chapters)?

  9. Generation architecture: Should images generate in parallel with text? Sequentially after? Pre-generate for each cast member as a "character sheet"? What's the flow that minimizes latency?

  10. Style choices: What visual style is both cheap to generate consistently AND appealing for a corporate-drama game (e.g., "boardroom thriller" aesthetic)? Think The Wire meets a graphic novel.

  11. Alternative lighter-weight options: Is there an 80/20 solution? Like just generating character portraits once per campaign (for the master doc cast) and showing them during dialogue blocks, without trying to illustrate every scene?

What I'm NOT Asking For

  • A full implementation plan or code
  • A specific vendor recommendation without cost analysis
  • Photorealistic character generation (too expensive, hits uncanny valley)
  • Animated or video output (way out of scope and budget)

What I AM Asking For

A thorough exploration that ends with: (a) a recommended architecture, (b) realistic cost per preamble, (c) the character consistency approach that will actually work, (d) an 80/20 option if the full comic-strip approach is too expensive, (e) a concrete "let's prototype this next" path.

Take your time. I want to understand the real tradeoffs before I spend time building anything.