huggingface-vision-trainer

To train vision models on Hugging Face Jobs, use the provided production-ready scripts for object detection, image classification, and SAM segmentation. You need to authenticate with HF_TOKEN in job secrets to launch the workflows and persist results to the Hub.

Yes, you can fine-tune object detection models including DETR, RT-DETR, and YOLOS on custom datasets. The skill provides dedicated training scripts that handle COCO-style dataset preparation, augmentation, and evaluation for these architectures.

Hugging Face Jobs supports SAM and SAM2 segmentation training through dedicated scripts. The skill handles SAM prompts, dataset preparation, and end-to-end training pipelines to fine-tune segmentation models on cloud GPUs.

The best way to estimate costs for cloud GPU vision training is using the built-in cost estimation features of the skill. It evaluates various hardware flavors and production-grade workflows before submitting the final Hugging Face Jobs training run.

You do not need to manually upload fine-tuned image classification models, as the skill includes Hub persistence. It automatically pushes your trained ViT and timm models directly to the Hugging Face Hub after the cloud GPU job completes.