ernie-research/FoR-T2I

Can Text-to-Image Models Draw from the Right Frame of Reference?

Left is not always image-left. If a person is facing the viewer, their left hand appears on the right side of the image. A text-to-image model can therefore render a plausible scene with the requested objects while still drawing the spatial relation from the wrong perspective.

FoR-T2I is the official benchmark release for Can Text-to-Image Models Draw from the Right Frame of Reference?. It evaluates whether T2I models can follow directional descriptions when the specified frame of reference differs from the camera view.

Each item pairs a frame-of-reference prompt (FoR_prompt) with a camera-view control prompt (Cam_prompt) that describes the same target layout. This paired design separates ordinary image-coordinate layout ability from the additional difficulty of using an oriented object's own left, right, front, or back. The release contains 1,200 prompt pairs across three difficulty levels, together with structured layout metadata and evaluation tools.

Across 22 recent T2I models, we observe a consistent drop from camera-view prompts to FoR prompts. We also include a training-free mitigation study based on prompt rewriting and VLM-gated visual feedback.

Results Preview

Dataset Summary

This repository provides prompts and structured layout metadata only; generated model images are not included.

Examples

Usage

from datasets import load_dataset

# After uploading this repository as a Hugging Face dataset repo:
dataset = load_dataset("ernie-research/FoR-T2I", split="test")
print(dataset[0]["FoR_prompt"])
print(dataset[0]["Cam_prompt"])

For local use without datasets:

import json
from pathlib import Path

path = Path("data/benchmark_v1_1200/FoR-T2I_1200_prompts.jsonl")
rows = [json.loads(line) for line in path.read_text().splitlines()]
print(len(rows))

Data Fields

Field	Description
id	Unique benchmark item identifier.
level	Difficulty level, one of L1, L2, or L3.
camera_view	Camera setting, either top_down or eye_level.
FoR_prompt	Prompt describing the target layout through the oriented anchor object's frame of reference.
Cam_prompt	Camera-view control prompt describing the same target layout with direction words from the viewer/image frame.
objects	Object roles used by the prompt, including anchor and target categories.
layout_result	Structured layout with coordinates, resolved orientations, and relation metadata.
layout_signature	Compact evaluation-oriented representation of the same layout.
target_anchor_relation	Relation of the target relative to the anchor's own orientation.
target_image_relation	Equivalent target placement in the camera/image frame.
object_orientation	Required anchor orientation.
orientation_rule	Orientation-link rule for L3 cases, or null otherwise.
split	Release split identifier.

Data Creation

Objects are drawn from existing text-to-image benchmark vocabularies and filtered for concrete, visually recognizable categories with usable orientation cues. Each item stores the resolved layout, camera view, anchor orientation, target relation, and paired prompt texts.

Evaluation

FoR-T2I can be evaluated with the scripts in the companion GitHub repository. The automatic evaluator uses open-vocabulary detection, segmentation fallback, depth cues, and VLM-based anchor-orientation verification. A prediction is counted as correct only when the required objects are detected, the target spatial relation is satisfied, and the anchor orientation is correct.

Expected generation-result rows follow this format:

{"id": "FoR-T2I-v1-000001", "prompt_kind": "FoR", "image_path": "/path/to/image.png", "status": "ok", "model": "model-name"}

prompt_kind must be either FoR or Cam; each benchmark item should have one generated image for each prompt kind.

Limitations

FoR-T2I is intended for evaluation, not training. Automatic scores depend on the selected detector, segmentation model, depth model, and VLM judge. Published results should report all generation and evaluation settings.

License

The FoR-T2I benchmark prompts and structured layout metadata are released under Creative Commons Attribution 4.0 International (CC BY 4.0). Code in the companion repository is released under the Apache License 2.0. Third-party models, checkpoints, and services remain under their own licenses and terms.

Citation

@misc{gu2026texttoimagemodelsdrawright,
  title={Can Text-to-Image Models Draw from the Right Frame of Reference?},
  author={Zheyuan Gu and Ruihang Li and Yong Huang and Yiqian Zhang and Xiangzhao Hao and Jiaxin Niu and Jiahao Hu and Zhenyu Zhang},
  year={2026},
  eprint={2608.03357},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2608.03357}
}