Omni-Image-Editor

Enables users to select arbitrary regions within an image and apply AI-driven inpainting to remove, replace, or regenerate content in those areas. The system uses deep learning models (likely diffusion-based or GAN architectures) to intelligently fill masked regions while maintaining semantic coherence with surrounding pixels. Region selection is performed through interactive canvas tools in the Gradio UI, with the selected mask passed to the backend inference pipeline for processing.

Provides a Gradio-based interactive canvas component where users draw or click to define regions of interest for editing operations. The system captures mouse/touch events, renders the mask overlay in real-time on the canvas, and converts the visual selection into a binary or soft-edge mask tensor that is passed to downstream processing pipelines. Supports brush-based drawing with adjustable brush size and eraser functionality for mask refinement.

Supports uploading and processing multiple images sequentially through a job queue system managed by HuggingFace Spaces infrastructure. Each image is processed through the inpainting pipeline in order, with results aggregated and made available for download. The system leverages Gradio's built-in queue management to handle concurrent requests and prevent server overload by serializing inference operations.

Provides a dropdown or selection interface allowing users to choose between different inpainting model architectures (e.g., Stable Diffusion inpainting, LaMa, or other open-source models) before processing. The backend dynamically loads the selected model from HuggingFace Model Hub and routes the inference request accordingly. This enables comparison of model outputs and selection based on quality/speed tradeoffs without redeploying the application.

Automatically detects input image resolution and format (JPEG, PNG, WebP), normalizes to a standard working resolution for inference (typically 512x512 or 768x768), and scales results back to original resolution. Handles aspect ratio preservation through padding or cropping strategies. Supports both upscaling and downscaling depending on input size, with configurable quality/speed tradeoffs.

Displays live progress indicators (percentage complete, estimated time remaining) during inference operations through Gradio's progress callback system. Allows users to cancel long-running inpainting operations mid-process, freeing GPU resources and returning control immediately. Progress updates are streamed from the backend to the frontend without blocking the UI.

Caches inpainting results based on a hash of the input image and mask, allowing identical editing requests to return cached results without re-running inference. Uses content-addressable storage where the cache key is derived from image content rather than request metadata, enabling deduplication across different users or sessions. Cache is stored in memory or on disk depending on Spaces instance configuration.