⚠️ Breaking Changes

• source_image moved from data to metadata. predict() now stores the original input image in detections.metadata["source_image"] instead of detections.data["source_image"]. This fixes IndexError when boolean- or integer-indexing sv.Detections with include_source_image=True (the default) — supervision indexes every value in data by the detection mask, but passes metadata through unchanged. Update any code that reads the source image: (#972)

  ```python

  Before (1.6.3–1.6.4)

  image = detections.data["source_image"]

  After (1.6.5+)

  image = detections.metadata["source_image"] ```

🔧 Fixed

• Fixed segmentation training crash on T4 and P100 GPUs. cuDNN engine selection fails for depthwise convolution backward on some CUDA stacks (Kaggle, Colab). The previous workaround only disabled cuDNN in the forward pass; backward kernels still ran with cuDNN enabled. A custom autograd.Function now disables cuDNN in both forward and backward. (#967)

• Fixed EMA segmentation mAP logged from base metrics. ema_segm_mAP_50_95 and ema_segm_mAP_50 were computed from the base (non-EMA) metric accumulator instead of the EMA accumulator, producing misleading validation scores for segmentation models. Both metrics now use the correct EMA source. (#980)

• Fixed BestModelCallback losing best EMA score on resume. The _best_ema value was not persisted in state_dict(), so resuming training reset the EMA best-model tracker and could overwrite a better earlier checkpoint. (#973)

• Fixed positional_encoding_size not updating with custom resolution. Setting resolution at construction time (e.g. RFDETRLarge(resolution=640)) did not update the positional encoding grid size, causing shape mismatches during forward. A model validator now auto-syncs positional_encoding_size when a custom resolution is provided. (#956)

• Fixed pretrained weight loading crash with custom resolution. Loading COCO pretrained weights into a model with a non-default resolution failed because the DINOv2 positional embeddings had an incompatible shape. Pretrained PE tensors are now bicubic-interpolated to match the target grid before loading. (#964)

  ```python

  This now works — PE is automatically interpolated from 560px grid to 640px

  model = RFDETRLarge(resolution=640) ```

• Fixed cryptic error on patch_size mismatch. When loading checkpoints without explicit args.patch_size, the first error was an opaque RuntimeError from load_state_dict. The compatibility check now infers patch_size from the DINOv2 projection weight shape and raises a descriptive ValueError explaining the mismatch. (#971)

• Fixed source_shape causing TypeError on sv.Detections iteration. predict() stored source_shape as a Python tuple, which sv.Detections cannot index per-detection. It is now an np.ndarray of shape (N, 2) with dtype int64, where each row is [height, width]. (#966)

• Fixed spurious "class_id out of range" warning for background class. RF-DETR uses num_classes + 1 logits internally; class index num_classes is the background/no-object class and is expected. Background-class detections now map class_name to "__background__" without a warning. (#970)

🏆 Contributors

Welcome to our new contributors, and thank you to everyone who helped with this release:

• Md Faruk Alam (@farukalamai) (LinkedIn)— cuDNN depthwise conv backward fix for T4/P100
• M. Fazri Nizar (@mfazrinizar) (LinkedIn) — EMA segmentation metrics fix
• Jirka Borovec (@Borda) (LinkedIn) — release coordination, reviews

Automated contributions: @copilot-swe-agent[bot], @pre-commit-ci[bot]

Full changelog: https://github.com/roboflow/rf-detr/compare/1.6.4...1.6.5