harpreetsahota204 / CRADIOv4

What is CRADIOv4?

This Python project implements C-RADIOv4, NVIDIA's latest vision foundation model, as a remote zoo model for FiftyOne. It lets you pull in these distilled models—blending SigLIP2, DINOv3, and SAM3 strengths—directly into your FiftyOne datasets for embedding extraction or spatial attention heatmaps. Register once via GitHub URL, then compute batched embeddings or PCA-based heatmaps on any image collection, scaling from 128px to 1152px+ with GPU acceleration.

Why is it gaining traction?

Unlike raw Hugging Face loads, it slots seamlessly into FiftyOne workflows: batch embeddings for similarity search, UMAP viz, duplicate detection, or representativeness scoring—all with one command like `dataset.compute_embeddings(model)`. Dual outputs (global summaries or spatial heatmaps with optional smoothing) and ViTDet mode for fast high-res inference make it a drop-in upgrade for dense perception tasks, competitive with DINOv3 at fewer params. Developers grab it for the zero-setup path to SOTA vision features in familiar tools.

Who should use this?

Computer vision engineers curating FiftyOne datasets for search, clustering, or debugging model attention. ML teams prototyping zero-shot classification, semantic segmentation probes, or heatmap viz on custom images. FiftyOne users implementing model zoo extensions for production pipelines.

Verdict

Solid starter for FiftyOne + C-RADIOv4 fans, with thorough docs and examples covering embeddings to brain-powered analysis. At 13 stars and 1.0% credibility, it's early-stage—test on small datasets first—but worth forking if you're implementing remote zoo models today.

(187 words)