New Release: Ultralytics v8.4.15

[glenn-jocher]

Ultralytics v8.4.15 is out!

Ultralytics v8.4.15 makes real-world image ingestion smoother (hello HEIC/HEIF + .heif with lazy decoding + improved EXIF handling) and ships important fixes for metrics correctness, exports, CI reliability, and docs. If you work with phone photos, multilingual labels, or edge exports, this one’s a practical upgrade

You can see the full release details in the Ultralytics v8.4.15 release page and the v8.4.15 full changelog comparison.

---

New Features (Highlights)

HEIC/HEIF image support (priority)

• Adds ultralytics.utils.patches.image_open and monkey-patches PIL.Image.open to enable HEIC/HEIF decoding only when needed via pi-heif (lazy fallback on first decode failure)
  • PR: ultralytics 8.4.15 HEIC/HEIF image support by @glenn-jocher
• Expands supported extensions to include .heif (alongside .heic)
• Removes HEIC-specific dataset loader code in favor of a unified image reading pathway

Results.save() usability improvement

• Adds create_dir (default True) to control automatic parent directory creation when saving outputs
  • PR: feat: improve Results.save() with pathlib and optional directory creation by @ShuaiLYU
  • Docs: Document Results.save() automatic directory creation by @raimbekovm

Better plotting for non‑Latin text

• Improves matplotlib font defaults (prioritizes Arial Unicode MS) and auto-registers .ttf fonts from the Ultralytics user config directory for stronger multilingual rendering

---

Important Fixes

Metrics correctness (SciPy matching)

• Fixes SciPy-based matching (match_predictions) to maximize IoU during assignment (previous behavior could select worst matches), affecting TP/FP assignment and mAP when use_scipy=True
  • PR: Fix scipy matching in match_predictions to maximize IoU by @Mr-Neutr0n

EXIF auto-rotation for list sources

• Applies ImageOps.exif_transpose() when loading from lists of paths/URLs, reducing sideways/rotated phone photo issues
  • PR: Apply EXIF transpose when list of paths is used as source by @Y-T-G

---

Export & Edge Updates (IMX / IMX500)

• Enables IMX export on Linux ARM64 (removes ARM64 block), improves dependency support and test gating
  • PR: Enable IMX export for ARM64 by @lakshanthad
• Adds segmentation as a supported IMX export/benchmark task
  • PR: Enable segmentation models in IMX benchmarks by @lakshanthad
• Logs ExecuTorch version during export for improved traceability
  • PR: fix: add ExecuTorch version during model export into logging by @onuralpszr

---

Dev / CI Reliability

• Adds PyTorch 2.8 / 2.9 tests to slow CI
  • PR: Add PyTorch 2.8 and 2.9 tests to slow CI by @Y-T-G
• Makes export tests less flaky by skipping TFLite export tests unless torch>=1.13 on Linux
  • PR: test: Add TFLite export requirement for Torch version 1.13 by @onuralpszr
• CI maintenance + workflow reliability improvements:
  • ci: move Conda Builds CI to monitor from conda-forge/ultralytics-feedstock to ultralytics repo by @onuralpszr
  • Added max concurrency 16 to link checker workflow for the HTTP/2 error by @banica-ultralytics
  • Banica ultralytics http2 changes undone by @banica-ultralytics

---

Docs & Guides (and YOLO26 momentum)

• Custom trainer docs + refinements:
  • Add guide on customizing the YOLO trainer by @raimbekovm
  • Simplify CustomSaveTrainer example in custom trainer guide by @raimbekovm
  • Add custom trainer guide cross-reference in train.md by @raimbekovm
  • docs: update hyperlink for hyperparameter tuning in custom trainer guide by @onuralpszr
• YOLO26 vs YOLO11 Raspberry Pi 5 charts added
  • PR: Add YOLO26 vs YOLO11 performance charts for Raspberry Pi 5 by @lakshanthad
  • If you’re starting fresh, Ultralytics YOLO recommends YOLO26 for new projects (smaller, faster, more accurate, end-to-end across all tasks). You can read more in the YOLO26 model docs.
• Platform documentation updates (annotation → training → deployment):
  • Update platform docs and Update Platform docs by @glenn-jocher
  • Platform Docs screenshots by @sergiuwaxmann
  • Update platform API docs and Platform Rate Limit docs by @glenn-jocher
  • Explore workflows in the Ultralytics Platform docs or jump straight into the Ultralytics Platform for annotation, training, deployment, and monitoring.
• Additional docs cleanups and fixes:
  • Fix typos and grammar errors in documentation by @him2him2
  • Fix broken YOLO26 documentation link in vision-eye guide by @raimbekovm
  • Fix docstrings and Delete orphan markdown files in build_reference.py by @glenn-jocher
  • Fix uv conflicts between export and solutions extras by @fcakyon
  • Fix: get the correct package name for metadata.distribution by @Laughing-q
  • docs: fix conda badge link in CI documentation by @onuralpszr
  • docs: remove unneeded tokenizer_ve documentation file by @onuralpszr
  • docs: remove non-existed dataset generation section from COCO12-Formats documentation by @onuralpszr
  • Add YOLOE agnostic_nms default behavior note and Document CUDA OOM auto-retry in batch-size settings by @raimbekovm
• New doc videos added
  • Add video to docs, Add video to docs, Add video to docs by @RizwanMunawar

---

Quick upgrade

pip install -U ultralytics

Minimal sanity check:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.predict("image.heic")

---

New Contributors

Welcome and thank you!

• @him2him2 — first contribution via PR #23617
• @banica-ultralytics — first contribution via PR #23690

---

Feedback wanted

Try v8.4.15 in your real pipelines (especially with HEIC/HEIF libraries, mixed phone imagery, multilingual plots, or use_scipy=True validation) and share results/issues in the repo. Your feedback helps keep Ultralytics YOLO solid for everyone in the community