Skip to content

AI turns documents or topics into real, native PowerPoint decks—with native shapes, transitions and animations, data-backed charts and tables on demand, audio narration from speaker notes, and support for your own .pptx templates. · by Hugo He

docs(roadmap): record the illustration deployment layer

hugohe387ca7bdccb27
diff --git a/docs/roadmap.md b/docs/roadmap.md
index f1eeddb2..0970ea51 100644
--- a/docs/roadmap.md
+++ b/docs/roadmap.md
@@ -64,7 +64,9 @@ The past two months' structural capability growth. Single flags / incremental po
- **A batch of default-behavior and intake standardizations** — per-element entrance animation now defaults off (page transition `fade` only; element builds are opt-in via `-a auto` / `animations.json`), removing the auto-cascade "AI deck" tell; per-project `icons/` copies chosen icons into the project at selection time and embeds project-first; `analysis/` is established as the canonical must-read layer for machine-extracted facts (PPTX intake bundle + `image_analysis.csv`); the main pipeline treats a source deck's identity (palette / fonts / layout) as **reference, not constraint** (inherit or redesign by Strategist judgment, defaulting to fresh design); the confirm page gains a custom color input
-- **AI illustration sheet → sliced spot-illustration pipeline** ([`scripts/slice_images.py`](../skills/ppt-master/scripts/slice_images.py)) — when a deck wants several spot illustrations of one family, instead of one AI call per cell it **generates a single multi-cell illustration sheet** (one call locks the whole set's style / palette and costs far less than per-cell generation), then `slice_images.py` deterministically cuts it on an `RxC` grid into individual element files; `--trim` tight-crops to each cell's content bounding box and `--alpha` knocks out the flat background, so each element lands as a **transparent cutout** on a differently-colored page rather than a visible box. Two §VIII resource rows make this explicit: one `ai` Illustration Sheet (generated but never placed directly, kept out of `spec_lock.md` images) plus one `slice` element row per cell (actually placed, listed in spec_lock); Step 5 generates the sheet, slices it, and re-runs `analyze_images`, and the Step 7 readiness gate handles the offline case by listing the sheet + element targets for the user to place and re-slice. The confirm page gains an `illustration_usage` binary intent toggle (`none` / `use`) — it states only *whether* illustrations join the deck's visual language; **how heavily and from which source are decided downstream** (intensity = Strategist judgment from content + `visual_style`; source = the existing `image_usage`, where `image_usage: none` always wins over illustration intent). Users never choose or see the internal sheet/slice implementation. `svg_quality_checker` gains matching validation, and [`image-layout-patterns.md`](../skills/ppt-master/references/image-layout-patterns.md) adds image-led promotional layout patterns
+- **AI illustration sheet → sliced spot-illustration pipeline** ([`scripts/slice_images.py`](../skills/ppt-master/scripts/slice_images.py)) — when a deck wants several spot illustrations of one family, instead of one AI call per cell it **generates a single multi-cell illustration sheet** (one call locks the whole set's style / palette and costs far less than per-cell generation), then `slice_images.py` deterministically cuts it on an `RxC` grid into individual element files; `--trim` tight-crops to each cell's content bounding box and `--alpha` knocks out the flat background, so each element lands as a **transparent cutout** on a differently-colored page rather than a visible box. Two §VIII resource rows make this explicit: one `ai` Illustration Sheet (generated but never placed directly, kept out of `spec_lock.md` images) plus one `slice` element row per cell (actually placed, listed in spec_lock); Step 5 generates the sheet, slices it, and re-runs `analyze_images`, and the Step 7 readiness gate handles the offline case by listing the sheet + element targets for the user to place and re-slice. Whether the deck uses spot illustrations is a Strategist judgment within the `image_usage` source boundary (not a separate confirm field; `image_usage: none` always wins over illustration intent); users never see the internal sheet/slice implementation. `svg_quality_checker` gains matching validation, and [`image-layout-patterns.md`](../skills/ppt-master/references/image-layout-patterns.md) adds image-led promotional layout patterns
+
+- **Spot illustrations: from "can be sliced" to "used well, as a system"** — a decision layer on top of the slice pipeline plus edge-quality hardening, so a deck actually uses illustration well rather than merely being able to generate it. ① **Slice quality**: `slice_images.py` `--alpha` now uses a feathered soft mask (anti-aliased + 1px defringe), samples the background from a 2px border-ring median, and measures color distance as the max channel difference; `svg_quality_checker` verifies generated `slice` rows have their files. ② **Trigger lean tied to `visual_style`**: each style carries a `core` / `supportive` / `sparse` illustration propensity ([`visual-styles/_index.md`](../skills/ppt-master/references/visual-styles/_index.md) `Illus.` column + each file's §6) — `core` styles recommend use by default, `sparse` default to none; precedence is `image_usage: none` → explicit user intent (overrides either way) → style propensity → none. ③ **Motif through-line**: when a deck leans into illustration, the cover anchor / section dividers / body spots are drawn from one motif family (shared h.5 rendering + palette + subject world) so the deck reads as one designed system rather than scattered spots; the AI motif generates only when `image_usage` includes `ai`, while `provided` / `web` reuse only assets that already form one family. ④ **Illustration-roles decision map**: a Strategist §h navigation table (role × when × mechanism × source — spot / hero anchor / section divider / atmospheric background / motif). Throughout: no quota, and no turning taste into a checker rule (a same-tiles placement detector was evaluated and declined, left to §4.3 placement prose + execution judgment)
- **Image transform-matrix fidelity end-to-end + host-native generation path** — `svg_to_pptx`'s DrawingML picture export now honors the SVG transform matrix (rotation / skew / compound transforms no longer shift or collapse to the origin under nested `<g transform>`), extending the cross-four-renderer position-fidelity axis to the raster layer; `image_gen.py` adds a host-native generation path that routes through the host's own image-generation capability when present. Both are fixes / incremental hardening — details in the commit log
@@ -75,7 +77,7 @@ The past two months' structural capability growth. Single flags / incremental po
Directions actively underway or up next, with no committed timeline.
- **Real-usage calibration of multi-deck intake and material divergence (just landed)** — multi-deck combined intake (`<stem>` prefixing + a `decks[]` merge index) and the material-divergence free-text field (under §c audience) both shipped (see "2026-06" above); next is calibration from real usage: same-stem collisions across decks currently let the later one overwrite the former — whether dedupe / numbering is needed awaits signal; and whether the Strategist reads the free-text intent accurately, and whether the "facts stay sourced" line holds when the user asks for free reshaping, awaits real generation. Neither gets a mechanical threshold added pre-emptively
-- **Real-deck calibration of the illustration-slice pipeline (just landed)** — the sheet → slice → place chain and the `illustration_usage` toggle shipped today (see "2026-06" above); next is calibration from real usage: whether one sheet sliced into many cells holds style / palette consistency in real generation, how robust `--trim` / `--alpha` are against irregular AI composition inside a cell, the offline readiness-gate manual-place + re-slice experience, and whether the Strategist turns `illustration_usage: use` into the right sparse-vs-recurring intensity. No mechanical threshold / quota added pre-emptively
+- **Real-deck calibration of the illustration capability (mechanism + deployment layer, just landed)** — the slice pipeline, the edge-quality hardening, and the decision layer (style propensity / motif through-line / roles map — see "2026-06" above) have all shipped; next is calibration from real usage: whether one sheet sliced into many cells holds style / palette consistency and how robust the `--alpha` soft mask is against irregular composition inside a cell, the offline readiness-gate manual-place + re-slice experience, whether propensity flips the styles it should, whether the motif reads as a designed system rather than over-decoration on real decks, and whether the source boundary holds (no silent AI on `provided` / `web`). No mechanical threshold / quota added pre-emptively; if same-size tiles really recur, consider a narrower lint then
- Otherwise: the mode / visual-style validation and calibration is settled (see "2026-06" above) — the structure (5 modes + 18 visual-styles + custom) is locked, the four adjacent pairs are consolidated into one Close-calls table, and the four calibration tightenings have landed. Future direction is driven by real usage signal and feedback; long-running improvements are under "Ongoing maintenance" below, and evaluated-out directions under "Non-goals"
---
diff --git a/docs/zh/roadmap.md b/docs/zh/roadmap.md
index 43bf4cd3..0cbfc9dc 100644
--- a/docs/zh/roadmap.md
+++ b/docs/zh/roadmap.md
@@ -64,7 +64,9 @@
- **一批默认行为与入口标准化** — 逐元素入场动画默认关(只留转场 `fade`;元素动画改 opt-in `-a auto` / `animations.json`),消除"自动级联入场"的 AI 味;per-project `icons/` 在选择时把选中图标复制进项目、嵌入优先本地;`analysis/` 确立为机器抽取事实的 canonical 必读层(PPTX intake bundle + `image_analysis.csv`);主管线把源 deck 的身份(配色 / 字体 / 版式)当**参考而非约束**(可继承可重设,由策略师判断,默认从零设计);confirm 页支持自定义配色输入
-- **AI 插画大图 → 切片点缀插画管线**([`scripts/slice_images.py`](../../skills/ppt-master/scripts/slice_images.py)) — 当一份 deck 需要若干同一家族的点缀插画时,不再一格一次 AI 调用,而是**一次生成一张多格插画大图**(单次调用锁住整组风格 / 色板、成本远低于逐格生成),再由 `slice_images.py` 确定性按 `RxC` 网格切成独立元素文件;`--trim` 按每格内容包围盒紧裁、`--alpha` 抠掉平整底色,让每个元素以**透明剪影**落到异色页面而非带可见方框。资源契约(§VIII)双行落地:一行 `ai` Illustration Sheet(生成但永不直接放置、不进 `spec_lock.md` images)+ 每格一行 `slice` 元素(实际放置、进 spec_lock);Step 5 生成大图后切片并重跑 `analyze_images`,Step 7 readiness gate 在离线场景列出大图 + 元素目标让用户手动放图再切。确认页加 `illustration_usage` 二元意图开关(`none` / `use`)——只表达「要不要让插画进入这套 deck 的视觉语言」,**用多重 vs 稀疏、走哪种来源都在下游决定**(强度=Strategist 按内容 + `visual_style` 判断;来源=沿用 `image_usage`,`image_usage: none` 永远压过插画意图);用户全程不必选、也看不到内部 sheet/slice 实现。`svg_quality_checker` 加了对应校验,[`image-layout-patterns.md`](../../skills/ppt-master/references/image-layout-patterns.md) 补了图主导的促销 / 宣传版式范式
+- **AI 插画大图 → 切片点缀插画管线**([`scripts/slice_images.py`](../../skills/ppt-master/scripts/slice_images.py)) — 当一份 deck 需要若干同一家族的点缀插画时,不再一格一次 AI 调用,而是**一次生成一张多格插画大图**(单次调用锁住整组风格 / 色板、成本远低于逐格生成),再由 `slice_images.py` 确定性按 `RxC` 网格切成独立元素文件;`--trim` 按每格内容包围盒紧裁、`--alpha` 抠掉平整底色,让每个元素以**透明剪影**落到异色页面而非带可见方框。资源契约(§VIII)双行落地:一行 `ai` Illustration Sheet(生成但永不直接放置、不进 `spec_lock.md` images)+ 每格一行 `slice` 元素(实际放置、进 spec_lock);Step 5 生成大图后切片并重跑 `analyze_images`,Step 7 readiness gate 在离线场景列出大图 + 元素目标让用户手动放图再切。要不要用点缀插画是 Strategist 在 `image_usage` source 边界内的判断(不单独成确认字段;`image_usage: none` 永远压过插画意图),用户看不到内部 sheet/slice 实现。`svg_quality_checker` 加了对应校验,[`image-layout-patterns.md`](../../skills/ppt-master/references/image-layout-patterns.md) 补了图主导的促销 / 宣传版式范式
+
+- **点缀插画:从「能切出来」到「会用、成体系」** — 在切片管线之上补齐决策层 + 收紧切片质量,让 deck 真正用好插画而非只是技术上能生成。①**切片质量**:`slice_images.py` 的 `--alpha` 改软蒙版抗锯齿 + 1px 去光晕、底色采样改 2px 边框环中位数、色距改最大通道差;`svg_quality_checker` 对 Generated `slice` 行校验文件存在性。②**触发倾向绑定 `visual_style`**:每个风格标 `core` / `supportive` / `sparse` 插画倾向([`visual-styles/_index.md`](../../skills/ppt-master/references/visual-styles/_index.md) 加 `Illus.` 列 + 各文件 §6),core 默认推荐用、sparse 默认不用;优先级链 `image_usage:none` → 用户显式意图(双向覆盖)→ 风格倾向 → none。③**贯穿母题 through-line**:deck 倾向用插画时,封面锚点 / 章节分隔 / 页内散点出自同一母题家族(共享 h.5 rendering+palette+主题世界),读成一套设计系统而非孤立散点;AI 母题仅在 `image_usage` 含 ai 时生成,provided/web 仅沿用本已成同族的素材。④**插画角色决策地图**:Strategist §h 加「角色 × 何时 × 机制 × source」导航表(散点 / 主角锚点 / 章节分隔 / 氛围背景 / 母题)。全程不设配额、不把品味数字化(同尺寸瓷砖检测评估后不做,留给 §4.3 placement 散文 + 执行判断)
- **图像变换矩阵端到端保真 + host-native 生成路径** — `svg_to_pptx` 的 DrawingML 图片导出现在如实尊重 SVG 的 transform 矩阵(旋转 / 斜切 / 复合变换不再在嵌套 `<g transform>` 下错位或塌回原点),把「跨四渲染器位置保真」主轴补到 raster 图层;`image_gen.py` 增加 host-native 生成路径,在宿主自带图像生成能力时走原生通道。两者均属修复 / 增量补强,细节见 commit log
@@ -75,7 +77,7 @@
明确在做或下一步要做的方向,不承诺时间窗口。
- **多 deck intake 与材料发散度的真实使用校准(刚落地)** — 多 deck 合并 intake(`<stem>` 前缀 + `decks[]` 合并索引)与材料发散度自由文字项(§c 受众下)均已上线(见上「2026-06」),接下来按真实使用信号校准:多份源 deck 同名(stem 冲突)的处理目前是后者覆盖前者,是否需要去重 / 加序号待信号;发散度的自由文字让 Strategist 判得准不准、放开写时「事实守源」边界守不守得住,待真实生成验证。两者都不预先加机械阈值
-- **插画切片管线的真实 deck 校准(刚落地)** — sheet→slice→place 整链与 `illustration_usage` 开关今日刚上线(见上「2026-06」),接下来按真实使用信号校准:一次大图切多格在真实生成里风格 / 色板一致性够不够稳、`--trim` / `--alpha` 对格内不规则 AI 构图的鲁棒性、离线 readiness gate 的手动放图 + 重切体验,以及 Strategist 把 `illustration_usage: use` 落成多重 vs 稀疏的判断准不准。不预先加机械阈值 / 配额
+- **插画能力(机制 + 部署层)的真实 deck 校准(刚落地)** — 切片管线、边缘质量收紧,以及决策层(风格倾向 / 贯穿母题 / 角色地图,见上「2026-06」)均已上线,接下来按真实使用信号校准:一次大图切多格的风格 / 色板一致性与 `--alpha` 软蒙版对格内不规则构图的鲁棒性、离线 readiness gate 的手动放图 + 重切体验、风格倾向是否翻对了该翻的风格、母题在真实 deck 上读成「设计系统」还是「过度装饰」、以及 source 边界(provided/web 不静默生成 AI)守得住否。不预先加机械阈值 / 配额;同尺寸瓷砖若真反复出现再考虑更窄的 lint
- 其余:mode / visual-style 体系的验证与校准已收口(见上「2026-06」),结构(5 mode + 18 visual-style + custom)定型、四对近邻消歧并成一张 Close-calls 表、四项校准收紧已落地。后续方向由真实使用信号与反馈驱动;长期改进见下「持续维护方向」,已评估不做的见「明确不做」
---