Project Files
docs / CHANGELOG.md
Notable changes to this project will be documented in this file.
w + h ≤ 1792 px). detect_object now operates exclusively on preview files instead of original files, eliminating OOM risk when processing large originals.mlx-vlm==0.1.13 package set. This fixes cannot import name 'load' from 'mlx_vlm' on /analyze.detect_object: crop percentages (cropLeft, cropRight, cropTop, cropBottom) stored per detection are no longer rounded to integers. Floating-point values are preserved so that downstream mask, , and tools receive the full sub-pixel precision when resolving a .detect_object (Qwen3-VL): default object-detection prompt configurable via plugin settings (Qwen3-VL: Object Detection Prompt).detect_object: improved tool description — backend-specific task syntax (Florence-2 tokens vs. Qwen3-VL natural language) now documented separately.detect_object: Qwen3-VL-8B supported as an alternative object-detection backend alongside Florence-2. Qwen3-VL produces more versatile detections and supports fine-grained labels and open-vocabulary prompts.detect_object: inference time text line now rendered for both the Florence-2 and Qwen3-VL backends.detect_object: canvas parameter replaced by targets (array or comma-separated string, 1–16 items, analogous to analyse_image). A single call now detects objects in multiple source images in one batch request to the Florence-2 /detect endpoint. Each source produces its own annotated output image (iN). Auto-select behaviour (omit targets when exactly one image is available) is preserved.cropzoom-indetectLabel