Skip to content

roadmap.md

Latest commit

 

History

History
48 lines (35 loc) · 1.8 KB

roadmap.md

File metadata and controls

48 lines (35 loc) · 1.8 KB
Raw

Roadmap / Backlog

Deliberately-deferred ideas, with enough context to pick them up later.

Reference-image role field (unify character / style / mixed)

Status: deferred — style transfer is intentionally kept prompt-based for now (Option C).

Problem

Today, character consistency has a structured-mode shortcut (subjectDescription → shared buildCharacterConsistencyPrompt in BaseTTIProvider), but style transfer has no equivalent — it is always prompt-driven. Reference images are generic and the prompt assigns the role (subject vs. style). This works (verified live with FLUX.2: robot + cat + a style ref in one request → both characters preserved and redrawn in the reference's art style), but offers no middleware-level convenience for style.

Proposed feature (Option B)

Add an optional role per reference image:

referenceImages: [
  { base64: robot, role: 'subject' },
  { base64: cat,   role: 'subject' },
  { base64: pixel, role: 'style' },   // role?: 'subject' | 'style' (default: 'subject')
]

The middleware would build the prompt from the roles (subject-consistency + style-transfer templates), covering character, style, and mixed cases in one unified API across providers.

Why it's a clean fit

Vertex Imagen already exposes native REFERENCE_TYPE_SUBJECT / REFERENCE_TYPE_STYLE. A role field could map onto those for Imagen, while Gemini and FLUX keep using prompt assembly.

Cost / scope

  • New optional field on TTIRequest → minor version (e.g. 1.15.0).
  • Prompt-builder changes in BaseTTIProvider (add a style template alongside the existing character template).
  • Per-provider wiring (Google Cloud, BFL).
  • Tests for character / style / mixed.
  • Backwards compatible (default role: 'subject' reproduces today's behavior).