SpotDiff is the first method to use a universal feature upsampler to rebuild self‑supervised vision‑transformer features at pixel resolution, a step that alone adds 13.8 F1 points over bilinear upsampling. The system stitches together three frozen foundation models: DINOv3 for semantic features, AnyUp for upsampling, and SAM 3 for text‑guided region proposals. Inside each proposed region the method compares the reconstructed pixel‑level features of two dates with cosine similarity; a fixed threshold (τ = 0.6 for object‑agnostic, τ = 0.7 for class‑specific) flags change. No learning or fine‑tuning is required.

On the LEVIR‑CD benchmark, SpotDiff with its largest ViT‑7B encoder reaches 82.46 % F1, which is 2.06 points higher than the best prior training‑free method and 14.46 points above the state‑of‑the‑art initialized approach. On the multi‑class SECOND benchmark its class‑average IoU climbs to 29.34 %, 1.74 % above OmniOVCD and 0.59 % above AdaptOVCD. When the same pipeline is run on PlanetScope (3 m) and Sentinel‑2 (10 m) imagery, it transfers unchanged, showing the same gains.

SpotDiff’s strength comes partly from the AnyUp upsampler. In an ablation on LEVIR‑CD, nearest‑neighbor upsampling produces 66.88 % F1, bilinear 68.69 % F1, while AnyUp reaches 82.46 % F1, confirming the upsampler’s crucial role.

The approach is limited by a few factors. It requires perfectly co‑registered RGB optical images; multispectral or SAR data are not yet supported. The fixed similarity threshold may hurt robustness across variable image conditions; an adaptive threshold could improve performance. Finally, the accuracy depends on the prompt supplied to SAM 3; different prompts can change F1 by up to 5.9 points.

Overall, SpotDiff demonstrates that composing modern foundation models can produce scalable, label‑free change detection that outperforms prior training‑free methods while avoiding costly annotation pipelines.