A robust OpenAI-compatible image-generation client with chat/images endpoint fallback, retry handling for transient errors, aspect-ratio validation and regeneration, and multimodal reference-image support. It is self-contained and could be extracted into any project needing reliable image-model invocation.

Uses a vision LLM to extract global style, theme colors/fonts, per-page layouts, JSON schemas, and external image slots from an uploaded PPTX render. Results are cached by content hash, and the module also handles layout assignment and prompt rendering, forming a reusable template-to-spec distillation pipeline.

Central normalization layer that converts heterogeneous inputs—cloned template profiles and sidecar style definitions—into a single validated RuntimeProfile contract with capability flags and prompt-strategy selection. Demonstrates a clean adapter pattern for unifying different style data sources behind one interface.

Provides deterministic image masking, compositing, and letterboxing utilities with strict size validation. It enables AI-edited image regions to be blended back into generated slides without corrupting the full visual, and is isolated enough to be reused in other image-generation or photo-editing pipelines.

A strict scene model that describes reconstructed PowerPoint slides as structured objects (text boxes, shapes, pictures, layers), enabling native editable PPTX output and versioned slide-level rollback. The validated structure is a reusable pattern for converting pixel-based generated documents into editable object graphs.