Image2Slides

Image-generated slide bases become source-grounded, editable PowerPoint decks. The workflow combines GPT-image-2 visuals, wiki planning, content boundaries, source-locked figures, and rendered-slide QA.

Preview of the reviewed howitworks Image2Slides deck
Human-reviewed PDF export from the howitworks example.
/image2slides

Slash-command surface

The plugin starts from required intent fields, writes the project wiki, and keeps each stage inspectable on disk.

GPT-image-2

Native image bases

Codex native image generation produces completed references and text-free backgrounds; source-locked figures are inserted later without redrawing results.

PPTX + PDF

Editable handoff

Backgrounds stay non-selectable, text remains editable, and the final reviewed PDF becomes the visual snapshot.

Workflow

From required inputs to a reviewed deck.

The v1 pipeline keeps the two GPT-image-2 batches explicit: completed references first, text-free background edits second, then deterministic PowerPoint assembly and boundary-aware QA.

Image2Slides workflow diagram
1Require tone, ratio, count, purpose, scene, and knowledge base before artifacts are created.
2Write the wiki, source registry, slide plan, and grep/generate content boundary.
3Generate GPT-image-2 completed slide references with visible text.
4Edit each completed image into a matching text-free background with identical geometry.
5Source-lock exact figures into detected panels and mark those panels non-editable.
6Analyze text/blank regions, build editable PowerPoint text over backgrounds.
7Render, run pixel/patch comparison, audit blank-zone boundaries, then export the reviewed PDF.

Commands

The CLI owns deterministic stages.

Agents make creative and factual decisions; the helper CLI keeps the file layout, composition, PPTX build, and QA repeatable.

/image2slides
style: white classroom report
ratio: 16:9
count: 5
purpose: presentation
scene: classroom
knowledge_base: ./howitworks

image2slides init --project decks/my-deck --spec spec.json
image2slides queue --project decks/my-deck

# after Codex native image_gen saves completed refs:
image2slides register-completed --project decks/my-deck

# after GPT-image-2 edits text-free backgrounds:
image2slides register-background --project decks/my-deck

image2slides compose-source-locked --project decks/my-deck
image2slides audit-layout --project decks/my-deck --strict
image2slides analyze --project decks/my-deck
image2slides build-pptx --project decks/my-deck
image2slides qa --project decks/my-deck --strict

# optional stricter blank-zone gate:
image2slides qa --project decks/my-deck --boundary-strict

Wiki

Every deck becomes an inspectable project.

The wiki is the plugin's working memory: intent, hierarchy, source boundaries, source records, and slide plans.

Project brief

wiki/00_project_brief.md captures the required inputs and keeps them out of visible slide text.

Content boundary

wiki/02_content_boundary.md states which claims must be sourced and which parts may be generated.

Source registry

wiki/03_source_registry.yml records documents, figures, extracted text, citations, and immutable data layers.

Slide plan

wiki/04_slide_plan.json defines titles, editable text, speaker notes, source layers, panel boxes, margins, and non-editable source-panel boundaries.

Content boundary

Factual material and generated scaffolding are not mixed casually.

grep_required

Facts, data values, citations, dates, paper names, quotes, screenshots, charts, and source figures. These must come from user material, web/search, or registered sources.

  • Source chart images remain exact.
  • Data/results are not redrawn by GPT-image-2.
  • Claims point back to the source registry.
  • Source panels are forbidden zones for editable text.

generation_allowed

Visual rhythm, teaching transitions, metaphors, non-factual phrasing, layout pacing, background motifs, and presentation polish.

  • Can improve explanation flow.
  • Can design blank panels and visual hierarchy.
  • Cannot invent unsupported results.

Input and output

A fixed contract keeps the workflow reliable.

Required inputs

Base style/tone, aspect ratio, slide count, purpose, scene, and knowledge base. The plugin does not treat these as optional.

Intermediate outputs

Prompt queues, native image bases, source-locked completed/background images, analysis JSON, and QA reports.

Final outputs

Editable pptx/image2slides.pptx and a human-reviewed pptx/image2slides.pdf visual snapshot.

Howitworks example

A minimal mental model committed with the repository.

The example uses the user's converted input PDF, extracted knowledge, GPT-image-2 bases, and reviewed PowerPoint/PDF result.

Preview pages from the howitworks input PDF

Input material

The input side is the converted howitworks.pdf, with extracted text and figures kept in the example folder.

Open input PDF
Preview pages from the reviewed Image2Slides output PDF

Reviewed output

The output side is the accepted PowerPoint export. This PDF is the canonical visual artifact for avoiding conversion drift.

Open reviewed PDF