モデルタイプ別のトレーニング解像度

トレーニング解像度はモデルの精度、推論速度、およびトレーニング時間に影響します。各モデルアーキテクチャにはこれらの要素のバランスを取るデフォルト解像度が設定されています。デフォルトでは、Roboflowは選択したモデルアーキテクチャに対するデフォルトのトレーニング解像度を推奨します。

下の表は各モデルアーキテクチャとサイズごとのデフォルトのトレーニング解像度を示しています。新しいデータセットバージョンを作成する際にリサイズ前処理ステップを設定することで、これらのデフォルトを上書きできます。 Dataset Version.

Object Detection

Model Type	Family & Size	Default Training Resolution
Object Detection	RF-DETR Nano	384×384
Object Detection	RF-DETR Small	512×512
Object Detection	RF-DETR Medium	576×576
Object Detection	RF-DETR Large	704×704
Object Detection	RF-DETR X Large	700x700
Object Detection	RF-DETR 2X Large	880x880
Object Detection	Roboflow 3.0 - Fast	640×640
Object Detection	Roboflow 3.0 - Accurate	640×640
Object Detection	Roboflow 3.0 - Medium	640×640
Object Detection	Roboflow 3.0 - Large	640×640
Object Detection	Roboflow 3.0 - Extra Large	640×640
Object Detection	YOLOv26(n/s/m/l/x)	640×640
Object Detection	YOLOv12 (n/s/m/l/x)	640×640
Object Detection	YOLOv11 (n/s/m/l/x)	640×640
Object Detection	YOLOv10 (n/s/m/b/l/x)	640×640
Object Detection	YOLOv9 (s/m/c/e)	640×640
Object Detection	YOLOv8 (n/s/m/l/x)	640×640
Object Detection	YOLOv5 (n/s/m/l/x)	640×640
Object Detection	YOLOv7 (legacy)	640×640
Object Detection	YOLO‑NAS Small	640×640
Object Detection	YOLO‑NAS Medium	640×640
Object Detection	Roboflow Instant	1008x1008

Instance Segmentation

Model Type	Family & Size	Default Training Resolution
Instance Segmentation	RF-DETR Seg Nano	312x312
Instance Segmentation	RF-DETR Seg Small	384x384
Instance Segmentation	RF-DETR Seg Medium	432x432
Instance Segmentation	RF-DETR Seg Large	504x504
Instance Segmentation	RF-DETR Seg X Large	624x624
Instance Segmentation	RF-DETR Seg 2X Large	768x768
Instance Segmentation	Roboflow 3.0 - Fast (Seg)	640×640
Instance Segmentation	Roboflow 3.0 - Accurate (Seg)	640×640
Instance Segmentation	Roboflow 3.0 - Medium (Seg)	640×640
Instance Segmentation	Roboflow 3.0 - Large (Seg)	640×640
Instance Segmentation	Roboflow 3.0 - Extra Large (Seg)	640×640
Instance Segmentation	YOLO-seg (v8/10/11/12)	640×640
Instance Segmentation	SAM3 (Segment Anything 3)	1008x1008
Instance Segmentation	Semantic segmentation (DeepLabV3+)	≥ 512×512

Classification & Pose

Model Type	Family & Size	Default Training Resolution
Classification & Pose	Resnet-18/34/50	224x224
Classification & Pose	YOLO-cls (v8/11)	224x224
Classification & Pose	Vision Transformer (ViT)	224x224
Classification & Pose	YOLO-pose (keypoints)	640x640

Multimodal/VLM

Model Type	Family & Size	Default Training Resolution
Multimodal/VLM	PaliGemma 2 - 3 B	448x448
Multimodal/VLM	PaliGemma 2 - 10 B/28 B	448x448
Multimodal/VLM	Florence-2	448x448
Multimodal/VLM	QWEN 2.5 VL	448x448
Multimodal/VLM	SmolVLM2	384x384