SUV / cat / bird
Cat above the SUV, bird anchored to the left side of the image frame.
ERNIE Team work
FoR-T2I
Can T2I Models Draw From the Right Frame of Reference?
Baidu ERNIE Team
1200
paired prompts
3
difficulty levels
256
coordinate frame
2
prompt views
Benchmark design
Object-centered relations, camera-centered controls.
Each sample pairs a FoR prompt with a camera-frame prompt that targets the same layout. The release stores structured coordinates, relations, orientation constraints, and prompt text in one evaluation-ready JSONL.
Cat / elephant / zebra
Elephant below the cat, zebra placed on the left side of the frame.
Sofa bed / TV / desk
TV right of the sofa bed, desk below the TV in a top-down layout.
Child / tiger / train
Three-object chain with a clear left-to-right viewer-frame ordering.
Van / woman / cat
Woman right of the van, cat above the woman in the image frame.
Red panda / polar bear / sheep
Eye-level depth sample with the sheep closer than the polar bear.
Release split
Balanced by level, explicit about camera and layout coverage.
Evaluation pipeline
Geometry first, orientation second, final paired score.
Build Manifest
Expand 1200 samples into FoR and Cam image rows with expected objects and relations.
Detect Objects
Use OWL-ViT with SAM fallback to localize required entities before relation scoring.
Score Geometry
Check left-right, top-bottom, and depth relations against the structured target layout.
Judge Orientation
Crop anchor objects and verify facing direction with a VLM judge.
Detected case gallery
Nine passing cases across L1, L2, and L3.
Hover over any case to inspect union bounding boxes. Use the view controls to pin the bbox overlay, DA3 depth map, or SAM fallback view; Image resets the card.
OWL-ViT boxes
SAM fallback boxes
DA3 depth map
Project resources
Released by the Baidu ERNIE Team.
The benchmark package, code, and paper links are grouped here for reviewers and downstream evaluators.