Finish the Hugging Face Hub integration #1027

Tracking the remaining work to make LibreYOLO a complete Hugging Face Hub library. Pick up whatever can be done on a given day.

Context

• Add LibreYOLO to model libraries huggingface/huggingface.js#2760 (merged 2026-11-06) registered libreyolo as a Hub library: label, repo link and download counting. Wauplin said it takes 2-3 days to go live on hf.co. Until then our model pages show 0 downloads and no "Use this model" button.
• 1.7.3 (Release 1.7.3: load models by Hugging Face repo id on a plain install #1020) made LibreYOLO("owner/repo") work on a plain pip install libreyolo (huggingface_hub is a core dependency now) and fixed the 10 YOLO9/YOLOX repos that also hold a legacy lowercase checkpoint. Checked on a fresh venv on Python 3.10-3.14 (macOS arm64): LibreYOLO9t, LibreYOLOXs, LibreDFINEn, LibreRFDETRn and others load and detect correctly.
• Audit of the 464 repos in the LibreYOLO org: LibreYOLO("LibreYOLO/<repo>") works on 429. About 31 more are VLM, generative and 3D models (Moondream, Molmo, PRX, TRELLIS, LGM, TripoSR, CRM, ShowUI, SenseNova, FCOS3D, DetAny3D, SAM 3D Body, ReID-osnet...) that load through LibreVLM, LibreGen, LibreGround or their own class, not LibreYOLO().

To do

1. Check that #2760 is live on hf.co: libreyolo shows as a library on a model page and downloads are counted.
2. Library: make LibreYOLO("LibreYOLO/<repo>") on those ~31 repos fail fast with the right call, e.g. "load it with LibreVLM("moondream-3")", instead of "multiple checkpoints" or "no loadable checkpoint". A small map from repo id to usage in libreyolo/utils/hf_hub.py, checked at the top of resolve_hub_checkpoint, plus a test against each family's HF repos. Ship it in a release before step 4.
3. Library: LibreYOLO("LibreYOLO/librefacerec-l") downloads into ./LibreYOLO/ in the current folder instead of the HF cache (libreyolo/models/facerec/weights.py). After that, every other LibreYOLO/<repo> load from the same folder is affected.
4. Send the follow-up PR to huggingface/huggingface.js. It is prepared on EHxuban11/huggingface.js:libreyolo-snippet (three commits, 11 lines, lint/format/check/test pass):
   • "Use this model" snippet: from libreyolo import LibreYOLO / model = LibreYOLO("<repo_id>") / results = model("path/to/image.jpg").
   • countDownloads also matching model_index.json and pipeline.json: 13 repos (PRX, ImageDream, TRELLIS) currently match nothing and always count 0.
   • filter: true in its own commit, offering to drop it until download numbers are visible.
   • No custom tags to hide the snippet. The description mentions the ~31 non-LibreYOLO() repos and offers to handle them however the reviewer prefers; it says they point to the right call, which is only true once step 2 is released.
     Send it after step 1, and thank Wauplin in the thread.
5. Hub model cards (no PR needed, edits in the LibreYOLO org):
   • 22 repos have the wrong pipeline_tag: every -seg RF-DETR, D-FINE and RTMDet repo says object-detection instead of image-segmentation, LibreYOLO9s-pose says object-detection, and Gemma4, Moondream and Qwen3VLEmbedding are off too. 159 cards have no explicit pipeline_tag.
   • 392 cards still show the filename form LibreYOLO("LibreYOLOXs.pt"); add the repo-id form LibreYOLO("LibreYOLO/<repo>"). 16 cards have no usage example.
   • Optional cleanup: delete the legacy lowercase libreyolo9*.pt / libreyoloX*.pt duplicates in the 10 YOLO9/YOLOX repos. Only the 0.1.x releases (no longer on PyPI) used them.
6. Port the 1.7.3 and 1.7.4 fixes from release to dev.