Releases: wep21/assets
Release list
SCRFD (face detection)
SCRFD-10G face detector (scrfd_10g_bnkps) with 5-point keypoints. Source: fal/AuraFace-v1. Used by brachio's face filter.
- scrfd_10g_bnkps.onnx — fp32
- scrfd_10g_bnkps-fp16.onnx — fp16 (modelopt autocast, keep-io-types so I/O stay fp32)
SAM3-LiteText text encoders (ONNX)
MobileCLIP text encoders for SAM3-LiteText (vil-uob/sam3-litetext-{s0,s1,l}), ONNX fp32 + fp16 (modelopt AutoCast). Reuses the SAM3 image vision-encoder and geo-encoder-mask-decoder ONNX (jamjamjon/assets 'sam3'); only the text encoder is variant-specific. Drop-in for the SAM3 text-encoder: inputs input_ids[B,32], attention_mask[B,32]; outputs text_features[B,32,256], text_mask[B,32]. Verified ORT==PyTorch (fp32 cos=1.0, fp16 cos~=0.99999).
HRNet (DarkPose)
HRNet top-down pose estimation weights with DarkPose decoding (COCO body 17kpt & COCO-WholeBody 133kpt), w32/w48 backbones, 256x192/384x288 inputs. fp32 and fp16 variants. Naming follows the jamjamjon/usls assets convention.
EdgeCrafter ECDetSeg (instance segmentation)
EdgeCrafter ECSeg instance-segmentation models (COCO-80) exported to ONNX for usls. FP32 (ecdetseg-{s,m,l,x}.onnx) and FP16 (-fp16.onnx, fp32 IO via NVIDIA ModelOpt autocast). Inputs: images[N,3,640,640] + orig_target_sizes[N,2]; outputs: labels, boxes, scores, masks[.,300,160,160].