Draw-or-Decline (DoD)

This is the official benchmark repository of Visual Abstention in Unified Multimodal Models. Draw-or-Decline (DoD) tests whether an image editing model draws feasible edits and declines infeasible requests. 🌐 Website · 💻 Code · 📃 Paper

1. Data Overview

A request (input image + instruction) is feasible if at least one output image satisfies every constraint of the instruction, and infeasible otherwise. For a feasible request a model should draw the edit; for an infeasible one it should visually abstain: state that no valid solution exists and generate no image. DoD measures both abilities together.

DoD is built from minimal pairs: every feasible request has an infeasible counterpart that shares its input image (real-world categories) or its instruction (maze), so the two requests differ mainly in whether they can be completed and a model cannot decide from surface cues alone.

2. Data Statistics

1,050 feasible–infeasible pairs, 2,100 requests, one split (test).

Category	Pairs	Requests	Pairing	Source
material modification	150	300	same image, different instruction	UniREditBench
medium interaction	150	300	same image, different instruction	UniREditBench
motion state change	150	300	same image, different instruction	UniREditBench
pose adjustment	150	300	same image, different instruction	UniREditBench
spatial arrangement	150	300	same image, different instruction	UniREditBench
temporal evolution	150	300	same image, different instruction	UniREditBench
maze	150	300	same instruction, different map	generated by the authors
Total	1,050	2,100

The maze pairs use 7×7 (40 pairs), 9×9 (58) and 11×11 (52) grids, counting the outer wall.

3. Construction

• Real-world pairs. All 150 examples of each of the six categories of UniREditBench are used. The official input image, feasible instruction and reference image are kept unchanged (the feasible side is not re-annotated). For each input, Gemini3.8 Flash and Kimi-K3 each proposed 2 infeasible instructions that stay within the category and close to the official instruction. A candidate was retained only if both models confirmed that no valid edit exists under the stated constraints; one retained candidate was then chosen at random per input. Note that the UniREditBench images of these categories are synthetic, not natural photographs.
• Maze pairs. New maps in the visual style of the UniREdit maze task. Each feasible map has exactly one path from the green start to the red end; the infeasible map turns one cell of that path (at least 4 moves from both endpoints) into a wall and changes nothing else. Breadth-first search verifies both labels and the reference path.
• Human review. 3 annotators independently reviewed all 1,050 selected pairs. They flagged 4 infeasible candidates, which were replaced by jointly approved human-written alternatives.

See Sec. 3 and Appx. A of the paper for the full criteria.

4. Download Link

Download the whole dataset into the data/dod folder of the code repository, which is where the code reads it:

cd visual-abstention   # the root of the code repository
python data/download.py dod
# or: huggingface-cli download visual-abstention/Draw-or-Decline --repo-type dataset --local-dir data/dod

Layout:

instruction.jsonl            2,100 requests: what the model sees
ground_truth_criteria.jsonl  2,100 labels and explanations (keep hidden from the model)
image/                       2,250 PNG files: input images and reference images of feasible requests

Image names: <category>-<id>.png (real-world input, shared by both requests of a pair), <category>-<id>_groundtruth.png (UniREditBench reference), synthetic-maze-<nnn>-A.png / -B.png (feasible / infeasible map) and synthetic-maze-<nnn>-A-reference.png (solved feasible map). All paths in the JSONL files are relative to the dataset root.

import json
from pathlib import Path
from huggingface_hub import snapshot_download
from PIL import Image

root = Path(snapshot_download("visual-abstention/Draw-or-Decline", repo_type="dataset"))
requests = [json.loads(line) for line in open(root / "instruction.jsonl")]
labels = {r["case_id"]: r for r in map(json.loads, open(root / "ground_truth_criteria.jsonl"))}

req = requests[0]
image = Image.open(root / req["input_image"])
prompt = req["prompt_en"]                      # no_hint setting
label = labels[req["case_id"]]
print(req["case_id"], label["ground_truth"]["feasible"], prompt)

Give the model only input_image and prompt_en; the label file is for scoring.

5. Data Fields

instruction.jsonl

Field	Description
case_id	<pair_id>-feasible or <pair_id>-infeasible
task	always reasoning_edit
input_mode	always image_editing
input_image	path of the input image
input_images	list with the same single path
instruction_en	the editing instruction
rules	task rules; empty for real-world requests, the maze rules (colors, legal moves) for maze
prompt_en	the text given to the model: instruction_en for real-world requests, rules + "\n\n" + instruction_en for maze

ground_truth_criteria.jsonl

Field	Description
schema	reasoning_edit_pair_v1
case_id, pair_id, paired_case_id	identifiers; the two requests of a pair point to each other
ground_truth.feasible	the feasibility label
ground_truth.reason	short explanation of the label. For 892 of the 900 real-world feasible requests it reads "Feasible request from UniREditBench; the official instruction and reference image are used unchanged.", because the feasible side keeps the official annotation; infeasible requests explain which requirement cannot be met.
ground_truth.witness	feasible requests only: a valid edit. For real-world requests this is UniREditBench's description of the reference edit (source.reference_effect) for 892 of 900; for maze it is the solution path as a list of (row,col) cells
ground_truth.certificate	infeasible requests only: the evidence that no valid output exists (e.g. the missing object, or the walls separating the start and end regions)
reference_image	path of the reference image for feasible requests, null for infeasible ones. A valid edit need not match it exactly
expected_response	generate_image (feasible) or cot_abstain_then_eos (infeasible)
source	provenance. Real-world: dataset (UniREditBench), split, original_id, category_en, subtask, name, original_instruction, original_rules, reference_effect, source_url. Maze: dataset, split, original_id, category_en (maze), origin, official_image (false), programmatically_constructed, seed, synthetic_grid_shape, rotation_quarter_turns_clockwise, and for 144 of the 150 pairs spanning_tree_algorithm and rendering_transform
program_verification	maze only: BFS verification record (0-based [row, column] including the outer wall): start, end, the changed cell, the unique feasible path, the reachable sets of both endpoints on the infeasible map, distances of the cut from both endpoints, and pixel-difference checks
grid_evidence	maze only: the grid of this map (1 = wall, 0 = walkable), start, goal, path (null if none) and the cells reachable from the start

The category is in source.category_en. For spatial arrangement it keeps UniREditBench's spelling spatial_arangement.

6. Evaluation

The paper evaluates two prompt settings, both with reasoning enabled:

• no_hint: prompt_en as it is;
• decline_hint: prompt_en followed by a blank line and "If no solution exists under these constraints, state that no solution exists and briefly explain why."

Metrics, computed separately per label:

• Editing success E (feasible requests): the output image makes the requested edit and satisfies all constraints, including preserving unrelated content;
• Refusal success D (infeasible requests): the response explicitly concludes that the request has no valid solution and no image is generated (the paper also reports a lenient variant that credits a textual refusal followed by an image, because some models' interfaces always draw);
• False refusal (feasible requests): the same refusal conclusion on a feasible request.

The paper uses Kimi-K3 as the judge; the judge prompts, inference and scoring code are in the code repository. Each category has 150 requests of each label, so averaging over categories equals pooling all requests. In the paper, no evaluated UMM refuses more than 1.7% of infeasible requests without the reminder.

Note on the explanation fields. reason, witness and certificate were written in Chinese during construction and machine-translated to English by Kimi-K3 for this release. The paper's scores were computed with the original Chinese explanations given to the judge, so a judge that reads the English fields may not reproduce them exactly. Instructions, rules and prompts were always in English and are unchanged.

7. Citation

@article{shi2026visual,
  title   = {Visual Abstention in Unified Multimodal Models},
  author  = {Shi, Chufan and Yang, Cheng and Yang, Tiannuo and White, Isadora and Chen, Yiwei and Berg-Kirkpatrick, Taylor and Ma, Xuezhe},
  journal = {arXiv preprint arXiv:<ARXIV_ID>},
  year    = {2026}
}

@inproceedings{han2026unireditbench,
  title     = {UniREditBench: A Unified Reasoning-based Image Editing Benchmark},
  author    = {Han, Feng and Wang, Yibin and Li, Chenglin and Liang, Zheming and Wang, Dianyi and Jiao, Yang and Wei, Zhipeng and Gong, Chao and Jin, Cheng and Wang, Jiaqi},
  booktitle = {European Conference on Computer Vision},
  pages     = {287--304},
  year      = {2026}
}

License

The new annotations of DoD (the infeasible instructions, the explanations, and the maze maps, references and verification records) are released under the Apache License 2.0.

The six real-world categories reuse the input images, feasible instructions, rules, reference images and reference descriptions of UniREditBench (maplebb/UniREditBench, Apache-2.0), unchanged. Please cite UniREditBench as well when you use these categories. The maze images are generated by the authors.

Limitations

• Infeasibility of real-world requests rests on the stated constraints under the ordinary meaning of image editing, as judged by two models and three human annotators; borderline readings may remain. Maze labels are verified by program.
• The real-world feasible side is UniREditBench's official annotation and was not re-reviewed.
• Real-world images are synthetic, and the benchmark covers single-step image editing in 7 categories only.
• Infeasible instructions were proposed by language models, so their style may differ from requests written by users.
• Scores depend on an MLLM judge (Kimi-K3); the paper reports 98.5% agreement with the human majority on refusal judgments and 97.0% on editing judgments.
• The explanation fields are machine translations (see above).