ultralytics 8.4.204

v8.4.204 - Release 8.4.204 with High Resolution Tiling documentation (#27069)

on Python PyPI

🌟 Summary

Ultralytics 8.4.204 makes high-resolution tiling available across training, validation, and prediction, and adds improvements to FastSAM prompts, model exports, and dataset handling.

📊 Key Changes

• 🧩 High-resolution tiling: Use tile=True or a positive integer tile size to process large images in overlapping windows. Training crops labels to each tile; validation and prediction merge results back into the original image.
• 📚 Updated tiling documentation and examples: A new High Resolution Tiling guide covers usage, supported tasks, and performance considerations. The SAHI guide remains available with a deprecation notice directing users to the built-in tile option, and the video example now uses the standard YOLO CLI and Python API.
• 🎯 Better FastSAM text prompts: FastSAM now uses a DataComp-XL CLIP model that selects the correct mask more often, with similar speed and download size.
• ⚡ Faster LiteRT quantization: INT8 and w8a16 calibration now use a C++ profiling mode. In the reported benchmark, export time dropped from about 36 seconds to 11 seconds, with no measured validation accuracy regression.
• 🛠️ More dependable exports and calibration: INT8 calibration now supports NDJSON and Platform datasets. NCNN export reports models its runtime cannot load, and dependency fixes address issues affecting TensorFlow, RKNN, Axelera, and Xilinx exports.
• 🔧 Additional fixes: RT-DETR INT8 export avoids known failures on certain SM 12.1 GPUs; resumed training uses new augmentation randomness; and the YAML parser accepts spaces before the colon after quoted keys.

🎯 Purpose & Impact

• 🐜 Tiling can preserve small-object detail in large images without requiring a separate tiling package. It works across several vision tasks, but increases the number of model passes and may take more time and memory.
• 🚀 Export and calibration changes can make deployment workflows faster or more reliable, particularly on affected hardware and formats.
• 📖 Users can use the built-in tiling option throughout training, evaluation, and prediction, while the existing SAHI guide remains accessible with a deprecation notice.
• 📦 The release updates the package version to 8.4.204. The release PR primarily updates documentation and examples; the other merged PRs provide the additional runtime and export changes.

What's Changed

• Keep RT-DETR INT8 squeeze convs and attention in float on SM 12.1 (GB10, N1X) by @atakann in #27049
• Use profiler-based calibration for faster LiteRT INT8 and w8a16 export by @amanharshx in #27062
• Accept spaces before the colon after quoted YAML keys by @MohammadHijjawi97 in #27052
• Fail NCNN export when NCNN cannot load the converted model by @raimbekovm in #27056
• Use DataComp-XL CLIP ViT-B/32 for FastSAM text prompts by @Y-T-G in #27060
• Fix saving multispectral detection crops by @Vaishnavi220506 in #27054
• Remove unused dataset YAML bash download route by @glenn-jocher in #27065
• Reseed RNG on resume so resumed epochs do not replay earlier augmentations by @Y-T-G in #27064
• Reuse dataset conversion for INT8 calibration by @glenn-jocher in #27067
• Fix dependency setup in TensorFlow, RKNN, Axelera and Xilinx exports by @Y-T-G in #26859
• Add high resolution tiling for all vision tasks by @Laughing-q in #27059
• Release 8.4.204 with High Resolution Tiling documentation by @glenn-jocher in #27069

Full Changelog: v8.4.203...v8.4.204