👋 Hello @heisenburrger, thanks for starting this detailed discussion and for using Ultralytics YOLO for your Coral TPU deployment work 🚀

This is an automated response to help move things forward quickly — an Ultralytics engineer will also review your post and follow up with more specific guidance soon 🙂

In the meantime, to help us better understand your scenario (large images, aspect ratio differences, and Coral TPU 416-size requirement), please share as many of the following details as you can:

1. Which Ultralytics YOLO model you are using (e.g., YOLO11, YOLO26, etc.), and the exact ultralytics package version
2. Your training configuration (image size, batch size, augmentation settings, model and data YAMLs if possible)
3. A few representative example images from your dataset, including their annotations (especially where detections become poor when trained at 416)
4. Training logs/plots (loss curves, mAP, precision/recall) comparing:
   • Training at your original large image size
   • Training at 416
5. Any preprocessing or tiling steps you have already tried (for either training or inference)

We also recommend checking the core docs while you wait, as many similar topics around image sizes and training setups are covered there:

• General docs: Docs
• Python examples: Python
• CLI usage: CLI

If this discussion eventually reveals a training-related issue or limitation, it can be very helpful if you can narrow things down as much as possible to a small, clearly described configuration (for example, a subset of your data and a minimal training command). That makes it much easier for us to reason about behavior and reproduce any potential problems.

If at any point this turns into a 🐛 Bug Report (for example, if you see inconsistent behavior between image sizes or environments), please provide a minimum reproducible example so our team can investigate efficiently.

If this remains a custom training ❓ Question (which it sounds like for now), please continue to:

• Include dataset image examples (especially showing your 1512×2040 samples and any tiling you are considering)
• Share your training logs and metrics
• Verify you are following our Tips for Best Training Results, as these practices often impact performance when changing image sizes or deployment targets like Coral TPU

You’re also very welcome to discuss tiling strategies and deployment considerations with the community:

• Real-time chat: Discord 🎧
• In-depth discussions: Discourse
• Community threads and experiences: Subreddit

Upgrade

Before we dive deeper, please ensure you are on the latest ultralytics version (many training and export improvements land there), and that your environment is up to date:

Upgrade to the latest ultralytics package including all requirements in a Python>=3.8 environment with PyTorch>=1.8 to verify your issue is not already resolved in the latest version:

pip install -U ultralytics

Environments

YOLO may be run in any of the following up-to-date verified environments (with all dependencies including CUDA/CUDNN, Python and PyTorch preinstalled):

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Using one of these can help rule out environment-specific issues when comparing training at large image sizes versus 416 for Coral TPU deployment.

Status

If this badge is green, all Ultralytics CI tests are currently passing. CI tests verify correct operation of all YOLO Modes and Tasks on macOS, Windows, and Ubuntu every 24 hours and on every commit.