TR-HASH Vision v6 · 1M · COCO

Compact hierarchical token-routed object detection

Intermediate research checkpoint. Training has not reached its planned 100-epoch budget. The published weights are provided for inspection, reproducible evaluation and runtime integration; they are not the final v6 release.

TR-HASH Vision v6 is a compact hierarchical detector with hash-routed spatial experts, native multi-scale features, a lightweight PAN/P2 neck and separate one-to-many and one-to-one prediction paths.

Live ComfyUI demo

The public Docker Space loads the EMA checkpoint below and exposes the one-to-one NMS-free branch as a visual ComfyUI workflow. Upload an image, queue the graph and inspect both the annotated result and JSON detections.

Published checkpoint

Checkpoint	Epoch	Step	Parameters	Resolution
`epoch-34-step-021595`	34 completed	21,595	997,542	512–640 px training, 640 px evaluation

The checkpoint directory contains:

ema.safetensors — recommended full detector weights;
model.safetensors — non-EMA full detector weights;
ema_tower.safetensors — reusable EMA vision tower;
tower.safetensors — non-EMA vision tower;
config.json, class_names.json and separate NMS/NMS-free metrics.

These are inference/evaluation checkpoints. Exact training resume also requires optimizer, scheduler, data cursor and distributed RNG state, which are not included in this model repository.

Intermediate COCO val2017 metrics

Inference path	mAP50	mAP50-95	AP small	AP medium	AP large	Best F1
One-to-many + NMS	0.1487	0.0607	0.0234	0.0519	0.0724	0.3029
One-to-one NMS-free	0.0713	0.0347	0.0119	0.0345	0.0469	0.1906

The one-to-many branch is currently stronger. The NMS-free path remains an experimental production-oriented branch and should not be presented as having parity with NMS at this checkpoint.

Architecture

Component	Vision v6
Backbone	Hierarchical TR-HASH tower with native P3/P4/P5 features
Attention	Local shifted windows with flexible 2D positions
Routing	4 hash-routed experts, top-2 active per spatial identity
Neck	Lightweight additive PAN with P2 small-object path
Head	Decoupled quality/classification and LTRB/DFL regression
Assignment	Dynamic one-to-many assignment with STAL small-object supervision
End-to-end path	One-to-one branch for NMS-free inference
Optimization	MuSGD, EMA, strong augmentation and multi-resolution training

Download and inference

from pathlib import Path
from huggingface_hub import snapshot_download
from trhash import Vision

snapshot = Path(snapshot_download(
    "AETHORIA-AI/TR-HASH-Vision-v6-1M-COCO",
    allow_patterns="checkpoints/epoch-34-step-021595/*",
))
checkpoint = snapshot / "checkpoints" / "epoch-34-step-021595"

model = Vision(checkpoint, runtime="torch", device="mps")
result = model("image.jpg", confidence=0.23)
result.save("prediction.jpg")

Use device="cuda" on NVIDIA systems or device="cpu" for the portable fallback. The separate trhash runtime is the intended user-facing API; the Complexity Framework remains the research and training codebase.

Resources

Limitations

This is an intermediate research checkpoint under CC BY-NC 4.0. Training is incomplete and accuracy remains well below mature production detectors. Validate calibration, latency and failure modes on the intended domain before deployment.

Downloads last month: -; Downloads are not tracked for this model. How to track

AETHORIA-AI
/

TR-HASH-Vision-v6-1M-COCO