Convert slide images, screenshots, scanned PDFs, and image-only PPT/PPTX files into high-fidelity object-level editable PowerPoint, including semantic-region mixed reconstruction, measured flowcharts and knowledge graphs, native circle nodes and connectors, single-object thin and filled arrows, spea

Nature Image2PPT Use this directory as the complete runtime. Run deterministic actions only through: Use Python 3.10 or later with requirements.txt installed. When a dedicated environment exists, substitute <image2ppt root /.venv/bin/python on macOS/Linux or <image2ppt root /.venv/Scripts/python.exe on Windows for every python command below. Do not continue after a failed doctor ; install only the reported missing dependency, then rerun it. Do not discover or invoke another Skill, CLI, Prompt, Schema, module, or state machine. Read the local contracts progressively Always read references/workflow.md . Read references/runtime dependencies.md only for setup or doctor failures, and read references/ocr text hints contract.md only when choosing or troubleshooting OCR. Before writing a page manifest, read references/page decision tree.md and references/manifest schema.md . Add only the references needed by that page: structured or compound page: references/region decomposition.md and references/object routing.md ; arrows: references/manifest arrow extension.md ; raster assets or image backend work: references/assets provenance contract.md . Before accepting or delivering output, read references/qa contract.md . Preserve the single source of truth Treat page jobs.json as the only page state source. Treat each pages/page NNN/manifest.json as the only page content source. Treat deck manifest.json as the final assembly source. Use only prepare , run next/dispatch/record/reset/hints/finalize , and the page commands in the local CLI for stateful lifecycle operations. Keep semantic region evidence in manifest.json.image2ppt region decomposition . Never create a second job file, reconstruction plan, OCR normalizer, page controller, packager, or finalize path. Let supplemental QA report failures; never let it mutate lifecycle state. Keep every write inside its owner directory Page build, validation, hints, and QA may read and write only inside that page directory. Manifest paths, recorded assets, formulas, reports, previews, and out overrides must not use .. , symlinks, or absolute paths to escape it. The sole external input exception is an explicit image tool result supplied to image import or as process sheet asset sheet source ; it is copied into the page before becoming a build dependency. Run level manifests and final outputs must remain inside the prepared run directory. Finalization rebuilds into a same directory temporary file and publishes it atomically only after a successful build. Treat any boundary rejection as a hard failure; do not copy the rejected file back into scope and present it as runtime output. Preserve pre migration behavior Treat self containment as a path/import/entrypoint migration, not a redesign of reconstruction behavior. Generate each worker Prompt from the complete local base layer plus the preserved Image2PPT profile layer. Do not condense, reinterpret, or replace either layer. Prefer the previously validated visual strategy when several routes satisfy the contracts. Keep simple measured objects native and retain bounded complex assets wherever a native redraw would reduce fidelity. Never re author an accepted baseline page merely to prove runtime independence. Run the workflow Image backend selection Use builtin imagegen when the agent runtime exposes image gen.imagegen ; it is the preferred backend because the worker can inspect edit inputs and import the explicit local result. Use the CLI image contract only when the built in tool is unavailable, errors, cannot read an input, or returns no valid local output. A missing optional argument such as model, mask, size, quality, or output path never authorizes fallback. Record the actual producer and permitted fallback reason in imagegen jobs.json . The CLI image contract is provider neutral at the transport boundary. Select codex oauth only for GPT Image model ids. Select openai compatible api for any provider specific model whose endpoint implements the OpenAI Images compatible /images/generations and/or /images/edits schema. Do not infer the image backend from the task's language model. Use an explicit backend when provenance matters; auto uses Codex OAuth only for compatible GPT Image ids and otherwise selects the configured API without sending Codex OAuth credentials to third parties. 1. Preflight and OCR choice Use Baidu AI Studio PADDLE OCR TOKEN when configured. If it is absent, tell the user once that the local builtin ink fallback measures text geometry but does not recognize characters; offer the configuration path in references/ocr text hints contract.md . Respect an offline only choice. 2. Prepare one run To pin a configured third party provider/model for auditable provenance, prepare with image backend openai compatible api . The run contract records the exact IMAGE2PPT IMAGE MODEL from the active project config or environment; it does not substitute a GPT Image default merely because no model flag was passed. Use no text hints only when OCR processing is intentionally disabled. Regenerate hints without creating a new run when needed: 3. Advance and claim pages For exactly one page, claim it with local and reconstruct it in the current agent. For multiple pages, dispatch independent page workers up to the capacity in page jobs.json . Do not reset a live worker merely because it is slow. 4. Reconstruct and gate each page Plan a structured page as 3–5 semantic regions and route each region independently. Use measured compound diagrams: measure every node, relation, and protected anchor. Keep measurable circles, cards, straight/dashed relations, and simple connectors native. Use bounded transparent assets only for complex local subparts. Represent a thin arrow as one connector with its arrowhead on the same object. Represent a filled arrow as one Arrow AutoShape, with centered label text inside the same object. Never construct an ordinary arrow from a line plus triangle and never flatten a whole knowledge graph into one image. Write new page manifests with schema version: 2 . Use structured visual inventory items with explicit kind and representation values, and write a concrete quality evidence observation for every required quality check. Formula rendering is a hard gate: a missing engine, converter, or failed compile must keep the page failed unless the user explicitly approves that exact formula exception and the manifest records both user approved exception: true and a concrete approval note . The worker Prompt performs the deterministic sequence. Its final gates are: The evidence file must cover the current source/render hashes and every required check with a specific observation. visual review notes is optional context and cannot substitute for the evidence file. Record only after standard validation and the Image2PPT region, arrow, and rendered gates pass: Use the same run reset → dispatch → record lifecycle to repair rejected pages. 5. Finalize and revalidate the rebuilt deck When run next reports finalize , run: Finalize rebuilds from page manifests, preserves source speaker notes, validates the package, and writes the output recorded by deck manifest.json . Final QA reapplies manifest arrows, verifies arrow atomicity and compound structure, renders every slide, checks speaker note integrity, and writes final/image2ppt qa.json . Deliver Return the final PPTX path, standard final validation, and final/image2ppt qa.json . Report which complex visuals remain replaceable bitmap assets. Do not call the deck complete while any page/final gate is pending or failed.

nature-image2ppt