mayanksingal7/JamaisVu

Jamais Vu: Benchmarking MLLMs on Subtle Physical Changes in Real-World Domestic Scenes

CVPR 2026 Workshop on Subtle Visual Computing (SVC). Paper: https://openaccess.thecvf.com/content/CVPR2026W/SVC/html/Singal_Jamais_Vu_Benchmarking_MLLMs_on_Subtle_Physical_Changes_in_Real-World_CVPRW_2026_paper.html

Jamais Vu tests whether multimodal LLMs can detect subtle physical changes between two photographs of the same real-world scene. Changes were made by physically manipulating household objects and re-photographing; nothing is synthetic or edited.

• 72 scenes, 356 images, 272 evaluation pairs
• Environments: living_room (41), bedroom (12), kitchen (6), bathroom (6), hallway (5), closet (2)
• Change types: add_remove (20), rotation (19), displacement (14), no_change (8), swap (7), state (4)
• Subtlety rated 1 (easy) to 3 (hard)

Dual-camera protocol

Every scene is photographed from two camera positions, before and after the change:

Slot	Camera	Phase	Role
A1	A	before	reference image
A1prime	A	before	re-take of A1 (identical scene)
C1	A	after	change, same view as A1
A2	B	before	identical scene, camera moved
C2	B	after	change, shifted view

From these, each scene yields four pairs (two for no-change scenes):

pair_type	images	view	has_change	what it measures
same_control	(A1, A1prime)	same	false	false-positive rate, pixel-aligned
shifted_control	(A1, A2)	shifted	false	viewpoint-induced hallucination
same_change	(A1, C1)	same	true	detection, pixel-aligned
shifted_change	(A1, C2)	shifted	true	detection under viewpoint shift

Files

• data/*.jpg — all images, named {scene_id}_{environment}_{slot}.jpg
• data/metadata.jsonl — one row per image with scene labels
• pairs.jsonl — the 272 benchmark pairs (use this for evaluation)
• annotations.json — scene-level source of truth from the annotation tool

pairs.jsonl fields

pair_id, scene_id, environment, image_a, image_b, pair_type, view, has_change, change_type, subtlety, description, mcq_options, mcq_answer

mcq_options is the correct description plus two distractors (shuffled with a fixed seed); null for control pairs.

Loading

from datasets import load_dataset
imgs  = load_dataset("mayanksingal7/JamaisVu", "images", split="test")
pairs = load_dataset("mayanksingal7/JamaisVu", "pairs",  split="test")

by_name = {r["file_name"]: r["image"] for r in imgs}
for p in pairs:
    img_a, img_b = by_name[p["image_a"]], by_name[p["image_b"]]
    # ... query your model with (img_a, img_b), compare to p["has_change"]

Or download the raw files and read pairs.jsonl directly:

hf download mayanksingal7/JamaisVu --repo-type dataset --local-dir jamais_vu

Suggested metrics

Report F1 / precision / recall on has_change, not just accuracy — models split into yes-biased and no-biased regimes that accuracy hides. Break down by view (same vs shifted) and by change_type.

Citation

@inproceedings{singal2026jamais,
  title={Jamais Vu: Benchmarking MLLMs on Subtle Physical Changes in Real-World Domestic Scenes},
  author={Singal, Mayank},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={1385--1391},
  year={2026}
}