Open
Description
Describe the proposed feature and its relevance to inferencing
Add support to slice, segment, and categorize large images atlases.
Used to generate a typical style for anomaly detection.
Tagging can be used to improve image organization.
Context or use case
I want to divide my image atlases into individual objects grouped by category for game art and tagging purposes.
Proposed solution
| Step | Description |
|---|---|
| 1. Integrate IREE.gd | Install and configure IREE.gd with an accelerated backend. This setup enables SAM 2 model execution directly on the GPU, optimizing segmentation for speed and low-latency processing. |
| 2. Load Image & Generate Masks | Load SAM 2 model via IREE.gd and segment the atlas image. Define prompts (e.g., clicks, bounding boxes) to generate segmentation masks for individual tiles. |
| 3. Organize & Categorize Tiles | Store each mask as a tile with names based on position (e.g., "tile_x_y") or content (e.g., "tree_tile") as identified by SAM 2. |
| 4. Optimize with GPU Parallelism | Use IREE.gd's parallel processing to run multiple segmentation tasks simultaneously, enabling faster, real-time slicing for large texture atlases. |
| 5. Save Segments with Metadata | Save each tile with metadata, including positional information and SAM 2 labels, to support quick filtering, retrieval, and categorization in automated workflows. |
Additional information
No response