Editorial Vision Studio

Editorial Vision Studio

AI Creative Director for Visual Generation.

Philosophy: Do not decorate. Always interpret.

An extensible Editorial Design Engine: one decision pipeline, swappable model adapters. Every layer — intent, analysis, art direction, planning, recovery, compilation, review, iteration — is a separate module with a single job, and the contract between them is a model-agnostic spec.

Architecture: reference/architecture.md

When to Use

• User asks for AI image direction, image prompts, art direction, visual concepting, or prompt adaptation across models
• User uploads a photo and asks for photo-to-design, editorial poster, cover, zine, gallery print, campaign key visual, brand visual, product visual, or hero image
• User gives a theme only and wants a poster, social asset, zine, campaign, moodboard, or conceptual image
• User mentions low-contrast / gray photo recovery (Panter compensation)
• User specifies a style: Swiss, Kinfolk, MUJI, Brutalist, Wallpaper*, Purple, Apartamento, POPEYE
• User wants analysis → direction → prompt → image, not immediate generation
• User specifies model: gpt-image, flux, ideogram — or asks to reuse direction with a different model
• User wants a set, not an image: campaign at all sizes, a carousel, a multi-page deck
• User wants a second image that matches the first — series consistency, brand system, "同一套視覺"
• User names a preset, or wants a specific fixed look reproduced exactly

Architecture: Decision Engine + Model Adapters

DECISION ENGINE (fixed)          MODEL ADAPTER (swappable)
Intent → Style Gate                   VisionSpec / EditorialSpec
      → Analyzer             →      ↓
      → Visual Language      →   adapters/{model}.md
      → Art Direction        →      ↓
      → Planner              →   GenerationRequest → API
      → Recovery             →
      → Compiler → VisionSpec
                                 SHARED POST-LAYER
                                 Reviewer → Optimizer → Generate
                                          → Evaluator → Iteration
                                                             │
                                 one spec mutation,          │
                                 re-enter Compiler ◄─────────┘
                                 (max 3 passes)

• Decision Engine emits spec/editorial-spec.schema.md — pure visual logic, zero model syntax
• Model Adapter translates spec → prompt (adapters/registry.md)
• Switching GPT Image → Flux → Ideogram: reuse VisionSpec / EditorialSpec, re-run adapter only
• Every run emits a VisualManifest (spec/visual-manifest.schema.md) so any run is replayable

Core Pipeline

User Request
    ↓
Intent Engine          → [prompts/intent.md](prompts/intent.md)
    ↓                    (series? → [prompts/series.md](prompts/series.md) · memory? → [prompts/visual-memory.md](prompts/visual-memory.md))
Style Gate             → [prompts/style-gate.md](prompts/style-gate.md)  ← ask the user once, up front
    ↓
Visual Analyzer        → [prompts/analyzer.md](prompts/analyzer.md)  (skip if theme-only / prompt-only)
    ↓
Visual Language Engine → [prompts/visual-language.md](prompts/visual-language.md)
    ↓
Art Direction Engine   → [prompts/art-direction.md](prompts/art-direction.md)  ← 2–3 candidates, commit to one
    ↓
Visual Planner         → [prompts/planner.md](prompts/planner.md)
    ↓
Recovery Engine        → [prompts/recovery.md](prompts/recovery.md) + [recovery/](recovery/)
    ↓
VisionSpec             → [spec/editorial-spec.schema.md](spec/editorial-spec.schema.md)
    ↓
Model Adapter          → [adapters/registry.md](adapters/registry.md)  ← swappable
    ↓
Prompt Reviewer        → [prompts/reviewer.md](prompts/reviewer.md)  ← is it legal?
    ↓
Prompt Optimizer       → [prompts/optimizer.md](prompts/optimizer.md)  ← is it well said?
    ↓
Image Generation
    ↓
Quality Evaluator      → [prompts/evaluator.md](prompts/evaluator.md)  → quality vector
    ↓
Iteration Engine       → [prompts/iteration.md](prompts/iteration.md)  ← fix one layer, loop (max 3)
    ↓
VisualManifest         → [spec/visual-manifest.schema.md](spec/visual-manifest.schema.md)

Each layer does one job. Never analyze in Compiler. Never generate in Analyzer. Never re-decide in Iteration — mutate the spec and re-run the layer that owns the failure.

Quick routing: reference/decision-tree.md

Step 0: Intent Engine

Before analyzing pixels, resolve user goal → output family:

Also resolve the six intent dimensions — subject, purpose, audience, emotion, platform, aspect ratio. Infer them; ask only when a missing one is load-bearing. They are what every later layer is graded against.

Detect scope here, not later: series_id for a set (prompts/series.md), memory_id for a continuation (prompts/visual-memory.md).

Read prompts/intent.md. Reject mismatched formats (e.g. gallery print for TEDx campaign).

Step 0.5: Style Gate

The first question about how the image should look. Before analysing a single pixel, ask which look they want — a numbered menu of the compatible presets plus 「讓 AI 提案」. (Intent may have asked about purpose or platform first; that is a different question.)

Half a step because it computes nothing. It asks one question and turns the answer into a lock.

要什麼風格？

1. 米色明信片 —— 米白紙底，照片重畫成簡化色塊，大量留白。
2. 時代海報 —— 飽和油墨滿版，硬邊平面色塊，地名做成版面。
3. 紙雕明信片 —— 寫實照片，明信片上長出立體紙雕世界。1:1。
4. 自己描述 —— 你說想要的樣子，我拆解照片後照你的描述重畫。
5. 讓 AI 提案 —— 看過照片和用途後給你三個方向再選。

Nobody chooses a look twice unless they asked to see proposals. The gate emits a single field, style_gate.outcome, and Art Direction branches on that and nothing else:

It skips — outcome: commit, no menu — when a preset or style is already named, a memory or series is active, a direction was committed earlier this session, or the run is unattended. It skips to outcome: offer when fewer than two presets fit the intent, because a one-item menu is not a choice. Unattended runs pick the highest-fit preset with intent-only weights and say which in one line — never silently.

Presets declare intended_layouts; the gate offers only those that intersect intent.allowed_outputs.

Show looks, not internal vocabulary. Nobody outside this repo knows what Kinfolk means.

The gate settles ground and medium only. Layout, composition, abstraction, and copy stay with the Planner.

Read prompts/style-gate.md.

Step 1: Visual Analyzer

Produce structured Image Report with star ratings and Editorial Score (0–100).

Dimensions: subject, clarity, contrast, saturation, composition, negative space, geometry, texture, lighting, emotion.

Read prompts/analyzer.md.

Step 2: Visual Language Engine

Derive Visual Language first, then style/palette/layout — not the reverse.

Examples: Museum → Swiss or MUJI + fine serif; Quiet Human → Kinfolk + cream/sage; Indie Memory → Zine + riso anchor; Poster Graphic → Travel Poster + flat inks.

Read prompts/visual-language.md. User style: override skips auto-derivation but Reviewer still validates DNA fit.

Step 3: Art Direction Engine

Do not generate the first plausible reading of the brief. Draft competing directions, each with a one-sentence thesis and a named trade-off, score them for fit, commit to one, and keep the runner-up.

Branch on style_gate.outcome and nothing else. The gate already resolved every case — preset chosen, free text, 「你決定」, explicit style:, memory, series, unattended, too few presets — into one of two values:

The layer always runs, even on commit — every spec needs an art_direction block with a thesis.

Every candidate must set a different ground. Eight of the eleven style DNAs resolve to a light paper field, so candidates chosen on style alone come back as three shades of ivory. Beyond that, candidates must differ on at least two of: render mode, visual language, layout family, abstraction level, typography weight. Palette swaps are not directions.

The committed direction is a hard constraint on the Planner. Switching to the runner-up later re-runs Planner onward only — never the Analyzer.

Read prompts/art-direction.md.

Step 4: Editorial Planner

Decide layout, typography direction, abstraction level inside the committed direction — not the final prompt.

Key rules (full matrix in prompts/planner.md):

• Portrait + negative space >50% → Magazine Cover
• Architecture + strong geometry → Swiss Poster
• Landscape + quiet mood → Gallery Print
• Street + human story → Documentary Zine
• Food/object + minimal → Product Editorial

The Planner also fills direction.spatial_plan — projection, shared ground plane, light direction, arrangement. On a photo run it is a verbatim copy of image_report.spatial; on a theme-only run the Planner authors it from intent.subject. Spatial fields locked by an active preset or memory are applied over the result field by field — a locked projection: flat-elevation is an authored decision that overrides what the photograph shows. It is required either way: prompts/reviewer.md rejects a prompt with no named projection whether or not a photograph existed, and without the field those clauses have no provenance for prompts/optimizer.md to protect.

If user specifies style: kinfolk, load styles/kinfolk.md DNA.

Step 5: Recovery Engine

Apply only when Image Report flags weakness. Each recovery is one atomic fix — see recovery/.

Panter Mode: discard dull grays; boost warm to 75% / cool to 70% saturation; add 8% high-chroma anchor block; widen tonal separation and mark scale. Panter is a colour compensation and never adds texture on its own. See recovery/contrast.md.

Texture Permission — single source of truth: assets/texture.md. Three tiers: PRINT (riso/halftone/scan defects) is zine only; SURFACE (substrate character such as cotton paper) is allowed on CLEAN layouts whose style DNA rates Texture ★★★+; FLAT (zero texture words) covers the photo-abstract-diptych panel ground, interface-asset, the website-hero copy-safe area, the product-editorial background, and every page of a presentation-deck that contains a data page. Recoveries never raise a layout's tier.

Never redesign the entire image unless Editorial Score <50 (Concept Reconstruction).

Step 6: Prompt Compiler + Model Adapter

Phase 1: Assemble VisionSpec / EditorialSpec — read prompts/compiler.md

Phase 2: Route to adapter by target.model:

User: model: flux or "用 Flux 生成" → set adapter, do not re-analyze.

Same direction, different model: reuse VisionSpec / EditorialSpec, swap adapter only.

Step 7: Prompt Reviewer

Before generation, run conflict detection. Read prompts/reviewer.md.

Examples:

• Swiss grid + Kinfolk organic → reject or resolve
• MUJI + heavy typography → reject
• Brutalist + soft pastoral palette → warn

Auto-correct incompatible pairings.

Step 7.5: Prompt Optimizer

The Reviewer answers is this prompt legal. It cannot fix a prompt that is legal but weakly worded — the ground clause buried in sentence four, a spec number softened to "a small accent", the same instruction stated twice.

Half a step because it decides nothing. Every clause it emits must trace to a field the spec already holds; a clause with no spec field behind it is a rejection back to the Compiler, not an invention. It never sets a ground, style, subject, layout, palette, or type scale.

Seven ordered ops: provenance_strip · ground_first · concretize · bind_numbers · dedupe · route_negatives · compress. Most gain comes from concretize and bind_numbers — a model renders nouns and numbers, never an adjective about taste.

This does not break the "mutate the spec, never the prompt string" rule in prompts/iteration.md. That rule forbids hand patches after generation. The Optimizer is deterministic, pre-generation, and recorded: same spec + same adapter + same ruleset_version produces the same string, so the manifest still replays.

No op firing is the normal outcome. The same op firing twice means the adapter emits that defect every time — fix adapters/{model}.md, not the string.

Variant mode emits 2–3 prompts from one spec, differing only on clause_density, specificity, and emphasis_order. A variant that differs on any spec field is a second direction, not a variant. Offer it only when the user asks to compare wordings, when the model is generic, or when a prompt-attributed dimension failed last run.

Read prompts/optimizer.md.

Step 8: Quality Evaluator

After generation, score a quality vector — not one number. Ten weighted dimensions: subject, composition, focal_point, palette, typography, texture, style_coherence, photo_fidelity, intent_fit, platform_fit. Each 0.00–1.00; inapplicable dimensions are null, never 0.

overall is the weighted mean and maps to grade A–D. Any single dimension below 0.60 fails the run regardless of overall.

The evaluator must name lowest_failing and responsible_layer — without them the iteration loop has nothing to act on.

Read prompts/evaluator.md.

Step 9: Iteration Engine

A low score points at one layer, not at the whole image. Fix that layer, recompile, regenerate.

lowest failing dimension → responsible layer → one minimal spec mutation → recompile → re-score

• One mutation per iteration. Two at once makes the next score unattributable.
• Mutate the spec, never the prompt string. A hand-patched prompt cannot be replayed from the manifest.
• Escalate when the same dimension fails twice: Compiler → Recovery → Planner → Art Direction → Intent.
• Re-enter at the Compiler after every mutation. "Responsible layer" names whose decision changed, not an entry point.
• Stop at overall ≥ 0.85 with no dimension below 0.60, at 3 iterations, on score regression, or when escalation reaches Intent — then report, don't guess.

Taste disagreement ("I don't like the blue") is a direction change, not a QC failure — route it to prompts/art-direction.md.

Read prompts/iteration.md.