image-svg-pptx-pro

# Image → SVG → Editable PPTX Pro Skill Use this skill when the user wants to turn a raster image, PPT screenshot, academic figure, UI screen, poster, report page, or AI-generated PPT image into a PowerPoint file that is as editable as possible while preserving visual fidelity. The core route is: `source image → normalized image → semantic layout plan → high-fidelity SVG → editable PPTX reconstruction → QA comparison → correction pass` This skill is intentionally **not** a naive OCR-to-PPT workflow. It uses SVG as the stable intermediate representation because SVG can encode precise geometry, typography, image crops, vector primitives, and fallbacks before converting to Office shapes. ## Output targets Always produce, when possible: 1. `reconstructed.pptx` — main deliverable; editable PowerPoint. 2. `reconstruction.svg` — high-fidelity intermediate source. 3. `layout_plan.json` — editable semantic layout plan. 4. `assets/` — cropped image/logo/icon/screenshot regions. 5. `qa_report.md` — visual and editability assessment. 6. Optional `macros/insert_svg_fallback.bas` — VBA helper for inserting SVG or fallback objects. ## Operating principle Prioritize final usefulness, not theoretical editability. - Convert text, headings, cards, boxes, dividers, simple charts, arrows, bullets, tables, and simple icons into editable PowerPoint elements. - Preserve complex photos, dense screenshots, logos, decorative illustrations, complex icons, gradients, and complicated path art as cropped image assets or SVG fallback groups. - Never redraw a complex logo/icon if it will visibly degrade. Crop it from the source image and position it precisely. - Never stretch text horizontally to match a screenshot. Use font size, letter spacing approximation, box width, and line breaks instead. - For Chinese slides, prefer legibility and correct text over forced font matching. ## Quality modes When the user does not specify a mode, use `balanced`. ### balanced Best default. Core text and structure editable; complex visual regions preserved as cropped assets. ### max_editable Use when the user explicitly wants more editable elements. Vectorize more lines, charts, and simple icons. Still keep complex logos/photos as assets. ### visual_locked Use when fidelity is more important than editing. Put a full-slide image/SVG fallback underneath or as grouped locked visual, then overlay editable text and key elements. ## Workflow ### 1. Normalize the source Run: ```bash python scripts/preprocess_image.py input.png --out work/normalized.png --meta work/source_meta.json ``` If the input is a phone photo of a screen, deskew/crop it first if needed. Keep the original file unchanged. ### 2. Build `layout_plan.json` Use visual reasoning to decompose the normalized image into semantic layers. Follow `references/layout_plan_schema.md`. The plan must include: - canvas width and height in pixels. - slide aspect ratio. - background color or background image. - text elements with exact text, x/y/w/h, font size, weight, color, alignment, line breaks. - vector elements: rectangles, rounded rectangles, lines, arrows, circles, charts, tables. - asset crop regions: photos, logos, screenshots, complex icons, decorative art. - z-order. - editability intent: `editable`, `asset`, or `fallback`. Do not invent unreadable text. Mark uncertain text as `needs_review: true`. ### 3. Crop assets Run: ```bash python scripts/crop_assets.py work/normalized.png work/layout_plan.json --out work/assets ``` Complex image regions referenced in the plan must point to these cropped files. ### 4. Generate SVG Run: ```bash python scripts/plan_to_svg.py work/layout_plan.json --assets work/assets --out work/reconstruction.svg ``` The SVG is the canonical visual intermediate. Inspect or render it before generating PPTX. ### 5. Convert SVG/plan into editable PPTX Prefer plan-based PPTX generation, because it preserves semantics better than arbitrary SVG parsing. ```bash python scripts/plan_to_pptx.py work/layout_plan.json --assets work/assets --out work/reconstructed.pptx ``` If the plan is unavailable but an SVG exists, use the SVG parser fallback: ```bash python scripts/svg_to_pptx_editable.py work/reconstruction.svg --out work/reconstructed.pptx ``` ### 6. QA and correction pass Create `qa_report.md` using `references/quality_checklist.md`. Check: - title hierarchy and position. - text correctness, especially Chinese characters and numbers. - margins and card alignment. - colors and background balance. - line thickness and rounded corner radius. - whether logos/icons degraded. - whether generated PPTX is actually editable for the main elements. If the visual mismatch is obvious, update `layout_plan.json` and regenerate. Do not present the first broken result as final. ## Recommended user-facing command When the user gives an image and asks for conversion, proceed without asking follow-up unless required files are missing. Say briefly what files were produced and link the PPTX/ZIP. ## Failure handling If exact text is unreadable, do not hallucinate it. Use cropped image fallback for the region and flag it in `qa_report.md`. If SVG-to-PPTX direct conversion loses filters, masks, gradients, or paths, keep those regions as cropped assets and overlay editable text/shapes. If the user needs a fully editable corporate deck, tell them which elements remain image-based and why.