Skip to content

🎯 Model card — lucid-yolo detector (n scale)

Covers the detection model accepted at the Det-smoke tier for release 0.1.0. One card per task family; segmentation and oriented detection get their own at 0.2.0 and 0.3.0.

📇 Model details
Developed by Independent reproduction, single operator; see PROVENANCE.md
Model type Single-stage anchor-free object detector, DFL-free dual head
Scale n — 2.4M parameters, 5.4 GFLOPs at 640×640
Input RGB, letterboxed to 640×640, float32 in [0, 1]
Output Two decode paths from one forward: NMS-free end-to-end (decode/topk_e2e.py, 300 detections as [x1, y1, x2, y2, score, class]) and a one-to-many path with NMS (decode/nms_path.py)
Version 0.0.1.dev12 (checkpoint from run v8, 2026-08-06)
License Apache-2.0 (code and report). Weights derive from COCO 2017 — see Licensing below
Paper Methods per the YOLO26 paper, arXiv:2606.03748. Independent implementation; not affiliated with, endorsed by, or derived from Ultralytics or its codebase

The architecture is typed Python, not a config DSL (ADR-001): five scale rows (n/s/m/l/x) multiply depth, width and max-channels over one topology. Only the n scale has been trained.

🧭 Intended use

Intended. Reproduction research: verifying the paper's architectural and training claims, ablating its mechanisms, and serving as a readable from-scratch implementation of NMS-free detection. The e2e path is the interesting artifact — it exports without suppression ops.

Not intended. Production detection, safety-critical or rights-affecting decisions, surveillance, biometric identification, or any deployment where 25 mAP on 80 common object categories would be mistaken for a reliable perception system. This is a smoke-tier reproduction at the smallest scale, roughly half the accuracy of a well-trained detector of this family, and it has been evaluated on exactly one benchmark.

🗂️ Training data

COCO 2017 train2017 — 118,287 images, 80 categories, annotations under CC-BY-4.0 with images under their original Flickr terms. No other data; no pretraining (D2: no Objects365 initialization).

COCO's documented composition limits apply to this model directly: category frequencies are long-tailed, scenes are web photography skewed toward particular geographies and contexts, and the person category carries all the demographic imbalance of that source. Nothing here corrects for it, and no fairness evaluation across subgroups has been run.

🏋️ Training procedure
Recipe batch 128, 50 epochs (46,250 steps), MuSGD (w_muon 0.5 / w_sgd 0.5), lr 0.02 with 3-epoch warmup then linear decay to lr0 × 0.01, weight decay 5e-4, momentum 0.95
Loss dual-branch; per branch CIoU 7.5 / classification 0.5 / L1 6.0 (stride units), Progressive branch weight 0.8 → 0.1
Assignment TAL (alpha 1, beta 6) with STAL small-target surrogates; one-to-many topk 10, one-to-one 7→1
Augmentation mosaic p=1.0 (disabled for the final 10 epochs), fused affine+letterbox, size-aware mixup and copy-paste, HSV jitter, horizontal flip p=0.5
Precision / seed 16-mixed, seed 0
Hardware one RTX PRO 6000, ~8m45s/epoch, ~7.3 hours total
EMA decay 0.9999, tau 2000

w_muon and w_sgd are independent gains (A7), so w_muon = 0 is a supported SGD-only configuration — MuSGD._parameter_update branches on it explicitly and skips the Newton–Schulz iteration rather than computing and discarding it, and tests/optim/test_musgd.py pins the resulting update. Supported is all it is: no run in this project has ever trained a tier under it, so there is no measurement of what the Muon branch is worth on this architecture and this row should not be read as one arm of an ablation. Every tier above trained at 0.5 / 0.5.

📊 Evaluation

COCO val2017, all 5000 images, torchmetrics MeanAveragePrecision on the faster_coco_eval backend. Both decode paths from one forward per batch.

weights path mAP50-95 mAP50 mAP75 mAP_S mAP_M mAP_L
EMA NMS 25.30 38.12 26.94 11.86 27.09 33.88
EMA E2E 23.84 36.00 25.15 11.64 25.86 31.38
raw NMS 25.33 38.45 27.05 — — —
raw E2E 24.23 36.72 25.82 11.24 — —

Recall at 100 detections (E2E, EMA): 46.07. The NMS-free path costs 1.11 AP against the NMS path on raw weights, 1.46 on EMA.

No latency figures are published. Throughput is hardware-dependent and the project treats parameter/FLOP fidelity as the substitute claim (all five scales within ±2% params, ±5% FLOPs of [R1 Table 7]).

⚠️ Limitations

  • Accuracy. 25.3 mAP is a smoke-tier result at the smallest scale, ~50 epochs against the paper's 500–600. Expect frequent misses and confusions on anything but large, unambiguous, well-lit instances.
  • Small objects are weakest — 11.9 mAP_S against 33.9 mAP_L. STAL is implemented but its benefit is a Det-ablations trend claim, unmeasured here.
  • 80 COCO categories only. No open-vocabulary capability; anything outside those categories is either missed or misclassified as the nearest one.
  • 640×640 letterboxed. Untested at other resolutions or aspect extremes.
  • One benchmark, one seed, one scale, one device. No variance estimate across seeds, and no cross-dataset generalization measurement at all: the RF100-VL tier that would have supplied one (WP-074/075) was dropped on 2026-08-18 over an unresolved per-dataset licence question, so nothing replaces it and none is scheduled. Training has also never spanned more than a single accelerator. This task's metrics are torchmetrics metrics that do synchronise across ranks, so a multi-device run is not obstructed the way the oriented and keypoint tiers are — but unobstructed is an untested inheritance from Lightning, not a result this project has.
  • No robustness evaluation — weather, blur, occlusion, adversarial or distribution-shift behavior is entirely uncharacterized.
  • Not bitwise reproducible across platforms (A26): libm last-bit rounding differs across OS and architecture; the goldens assert structural metrics with tolerance, not byte hashes.
  • Determinism is weaker on MPS than the recipe reads. default_determinism (cli/train.py) resolves deterministic to "warn_only" whenever MPS is the auto-picked accelerator, because MPS ships no deterministic kernel for some backwards this model hits (index_put_with_accumulate in the assignment/loss backward); strict True would abort the step. CPU and CUDA keep strict determinism. An Apple-silicon run is therefore reproducible only up to those non-deterministic kernels, and says so in warnings rather than by failing.
  • The seed reproduces a run only at the same worker count. --data.num_workers 0 replays the augmentation stream byte for byte; with workers each one is re-seeded per worker and per epoch from the loader's own generator (WP-079). Both are fully determined by seed, but they are different streams — a run at 32 workers does not reproduce the same seed at 0 workers, or at 16.

🤝 Ethical considerations

A detector trained on COCO inherits COCO's person category, and any deployment that detects people carries surveillance and privacy implications this project has not assessed. The intended-use section is a boundary, not a disclaimer: the model has no evaluation supporting use on people, and none is planned.

Failure modes are unbounded in the sense that matters — no calibration study exists, so the confidence scores should not be read as probabilities, and the model gives no signal when it is operating outside its training distribution.

📜 Licensing and provenance

Code and this report: Apache-2.0. Weights derive from COCO 2017; no trained weights are published (D14).

Clean-room statement. No file, configuration, weight, or code fragment from any Ultralytics repository, package, documentation site, or released checkpoint was opened, downloaded, imported, or consulted at any point in this project's history. Every design input traces to a paper, its cited primary literature, a registered assumption, or — for diagnosis only, never copying — a permissively licensed independent implementation registered under D13/ADR-004. The audit trail is PROVENANCE.md; the standing prohibitions are AGENTS.md sec. 7.

🔖 Citation

This reproduction has no publication. Cite the paper it reproduces:

@article{yolo26,
  title  = {YOLO26},
  note   = {arXiv:2606.03748},
  year   = {2026}
}