Organizations: Renmin University of China, China · Huazhong Agricultural University, China · Huazhong University of Science and Technology, China · Jiangnan University, China · Nanchang University, China
We study a systematic bias in modern image generation models: the mention order of entities in text spuriously determines spatial layout and entity--role binding. We term this phenomenon Order-to-Space Bias (OTS) and show that it arises in both text-to-image and image-to-image generation, often overriding grounded cues and causing incorrect layouts or swapped assignments. To quantify OTS, we introduce OTS-Bench, which isolates order effects with paired prompts differing only in entity order and evaluates models along two dimensions: homogenization and correctness. Experiments show that Order-to-Space Bias (OTS) is widespread in modern image generation models, and provide evidence that it is primarily data-driven and manifests during the early stages of layout formation. Motivated by this insight, we show that both targeted fine-tuning and early-stage intervention strategies can substantially reduce OTS, while preserving generation quality.
Figures & tables
Figure 1: Order-to-Space Bias makes models treat mention order as a cue for spatial layout and role binding. (a) With neutral prompts, the first-mentioned entity tends to appear on the left, and under-specified edits are often applied to a default side. (b) When order conflicts with grounded cues, the model follows order instead, producing incorrect layouts or role inversions.
Figure 2: Overview of OTS-Bench . We evaluate OTS in four settings: T2I and I2I, each with homogenization and correctness.
Component
Count
Source
Subjects
Humans
64
SOC 2018
Animals
25
COCO, ImageNet
Objects
49
COCO, OpenImages
Subjects Total
138
Actions & States
Table 1: OTS-Bench taxonomy overview. Subject and action/state inventories are curated with reference to COCO ( Lin et al., 2014 ) , ImageNet ( Deng et al., 2009 ) , OpenImages ( Kuznetsova et al., 2020 ) , AVA ( Gu et al., 2018 ) , VerbNet ( Schuler, 2005 ) , Animal Kingdom ( Ng et al., 2022 ) , and Something-Something ( Goyal et al., 2017 ) .
Modality
Homogenization
Correctness
Total
T2I
1,700
400
2,100
I2I
1,100
1,100
2,200
Total
2,800
1,500
4,300
Table 2: OTS-Bench summary.
Figure 3: Multi-agent pipeline for generating constrained test cases. Task Explorer proposes candidates, Insight Synthesizer validates real-world constraints, and Diversity Architect produces structured data.
Model
Homogenization
T2I Correctness (%)
I2I Correctness (%)
Cohen’s κ↑
T2I
I2I
Ali
Rev
Ali
Rev
llama-3.2-11b-vision
97.1
73.3
97.0
2.4
89.7
87.6
0.62
Qwen3-VL-8B-Instruct
94.2
63.9
85.3
12.5
90.4
83.7
0.81
InternVL3-8B
90.2
59.8
84.8
24.3
94.9
92.9
0.72
llava-critic-r1-7b
54.9
60.2
91.4
23.8
96.9
95.2
0.47
Human
90.3
68.7
82.0
9.2
89.8
82.5
–
Table 3: Alignment results between VL judges and human annotations. Homogenization : deviation score defined as Hom=∣2p−100∣ , where p is the percentage of valid samples assigned to label 1 (valid samples exclude label 0). Correctness : accuracy on valid samples under Ali and Rev . Cohen’s κ : computed on the full 2,400-sample 3-class confusion matrix (labels 0/1/2).
Model
Homogenization ↓
T2I Correctness ↑
Δ↓
I2I Correctness ↑
Δ↓
T2I
I2I
Ali
Rev
T2I
Ali
Rev
I2I
SDXL
52.6
55.8
83.3
23.6
59.7
73.5
62.2
11.3
SD3.5
84.2
38.4
84.8
21.1
63.7
79.5
71.8
7.7
FLUX-dev
88.8
42.6
79.8
24.7
55.1
84.3
76.8
7.5
Qwen-Image
91.6
83.4
81.8
28.2
53.6
88.1
78.8
9.3
DALL-E 3
70.4
–
87.7
18.2
69.5
–
–
–
Table 4: OTS-Bench main results. Homogenization measures deviation from a balanced 50/50 split in left–right outcomes: Hom=∣2p−100∣ where p is the percentage of valid samples assigned to the left (lower is better). Correctness reports accuracy when prompt order aligns ( Ali ) or contradicts ( Rev ) the grounding constraint. We report degradation Δ=Ali−Rev as a diagnostic of order sensitivity.
Dataset
Samples
OTS-align (%)
LAION-2B-en-aesthetic
23,020
89.2
DataComp-Large
10,150
87.2
Table 5: Order-to-space alignment in web-scale corpora. Samples is the number of valid caption–image pairs after filtering. OTS-align is the proportion of cases where the caption order (first entity, second entity) matches the image layout (left entity, right entity).
Model
Homogenization ↓
T2I Correctness ↑
I2I Correctness ↑
Image Quality ↑
T2I
I2I
Ali
Rev
Ali
Rev
ImageReward
FLUX-dev
88.8
42.6
79.8
24.7
84.3
76.8
0.186±0.672
FLUX-dev (SFT)
47.4
25.0
83.2
49.3
87.0
79.5
0.217±0.687
Qwen-Image
91.6
83.4
81.8
28.2
88.1
78.8
0.204±0.732
Qwen-Image (SFT)
50.8
32.8
84.7
57.7
90.5
80.1
0.223±0.717
Table 6: Merged comparison of homogenization, correctness, and image quality for FLUX-dev and Qwen-Image (with LoRA variants). Homogenization is lower-is-better ( ↓ ); correctness and image quality are higher-is-better ( ↑ ). Image quality is reported with ImageReward (mean ± std).
Switch step
Setting
base
1
2
3
4
5
(a) Switch direction: Successful Rate (%)
A2B
100.0
25.0
14.8
11.1
3.2
2.2
B2A
100.0
23.0
17.8
5.5
2.4
1.2
(b) Prompt transition: Homogenization (%)
N2A
89.9
28.8
21.8
27.2
34.1
39.1
Table 7: Temporal effects of order cues. (a) A2B/B2A: switching direction. (b) N2A/N2B: switching from neutral to rich prompts with different entity orders. Values are in %.
standing vs sitting, raising hand vs saluting, waving vs clapping, looking up vs down, leaning forward vs backward
Subtotal
35
Individual + interactive + parallel actions
Animal Actions
Static Postures
4
sitting, standing, lying down, crouching
Appendix
Table 9: Action library with comprehensive counts (unique actions; no double-counting).
Modality
Dimension
H–H
A–A
O–O
H–A
H–O
A–O
Total
T2I
Homogenization
500
300
300
200
200
200
1700
T2I
Correctness
–
–
–
–
–
–
400
I2I
Homogenization
500
300
300
–
–
–
1100
I2I
Correctness
500
300
300
–
–
–
1100
Total
4300
Appendix
Table 10: Full OTS-Bench breakdown by modality, evaluation dimension, and entity-pair type.
Agent Role
System Instruction
Prompt Scout
Objective: Generate prompts that expose sequence-to-spatial bias in text-to-image models. Critical Understanding: Sequence-to-spatial bias occurs when models incorrectly map text order to spatial positions (first→left, second→right). Valid test cases must have real-world layouts that contradict text order. Example: “gas pedal and brake pedal in a car”—reality has gas(right) and brake(left), but biased models generate gas(left) and brake(right) by following text order. Requirements: (1) Each prompt must test a verifiable spatial constraint (keyboard keys, anatomy, vehicle controls, etc.); (2) Elements A and B must have fixed positions in the given context; (3) Only left/right semantics allowed (no top/bottom, front/back); (4) Domain diversity: aim for 10+ distinct domains; (5) Total target: 400 test prompts. Deliverable: (1) Mission restatement confirming bias detection objective; (2) Draft prompt list with clear rationale for each domain; (3) Handoff notes for Constraint Critic highlighting coverage gaps. Archive requirement: label each round clearly for version control.
Constraint Critic
Objective: Review Prompt Scout’s draft and provide strict, objective feedback. Evaluation Criteria: (1) Spatial validity—does each A-B pair have a verifiable real-world left/right constraint? (2) Bias exposure—will the prompt actually trigger sequence-to-spatial errors? (3) Domain coverage—progress toward 10+ domains and 400 prompts; (4) Constraint adherence—only left/right axis used. Required Output: (1) Quantitative metrics (domain count, constraint coverage, total prompts, archive status); (2) Concrete strengths and weaknesses of current draft; (3) Numbered directives for Benchmark Curator; (4) Strict scores (0–5) for diversity, coverage, failure-exposure, and archive completeness. Critical: Flag any prompts where left/right positions are arbitrary or undefined.
Benchmark Curator
Objective: Rewrite benchmark based on Constraint Critic directives. Output Format: JSON per domain containing: domain name, round number, seed pairs with elements A/B, context, constraint type (must be “left_right”), semantic basis, difficulty level (obvious/subtle/tricky), spatial reality (A/B positions), rationale citing verifiable standards, sequence bias failure mode, test prompts with expected errors, and status (new/revised/retired). Critical Rules: (1) spatial_reality must cite verifiable standards (QWERTY layout, anatomy atlases, traffic codes, etc.); (2) Each test prompt must demonstrate how sequence bias causes spatial error; (3) Mark prompts as new/revised/retired vs previous round; (4) Concise format: “A and B in/on/at Context” (single sentence). Validation: Reject any pair where left/right positions are arbitrary or context-independent.
Appendix
Table 11: System instructions for the three-agent benchmark construction pipeline. Each agent receives role-specific prompts that enforce left-right constraint validation and bias detection objectives.
Setting
Value
Models
FLUX.1-dev, Qwen-Image
Trainable modules
UNet only (text encoder frozen)
LoRA (linear)
rank=32, alpha=32
LoRA (conv)
rank=16, alpha=16
Training steps
10,000
Batch size
1
Appendix
Table 12: Key hyperparameters for anti-OTS LoRA-SFT.
Qwen3-VL-8B-Instruct ( κ=0.81 )
llama-3.2-11b-vision ( κ=0.62 )
InternVL3-8B ( κ=0.72 )
llava-critic-r1-7b ( κ=0.47 )
Human
0
1
2
Total
0
1
2
Total
0
1
2
Total
0
1
2
Total
0
224
11
4
239
207
19
13
239
209
19
11
239
183
37
19
239
1
94
1475
30
1599
103
1377
119
1599
96
1471
32
1599
92
1342
165
1599
2
35
56
471
562
31
178
353
562
18
154
390
562
19
281
262
562
Total
353
1542
505
2400
341
1574
485
2400
323
1644
433
2400
294
1660
446
2400
Appendix
Table 13: Full 3-class confusion matrices (labels 0/1/2) between human annotations (rows) and VL judges (columns) on the 2,400-sample validation set. Each judge occupies a 4 × 4 block (including marginal totals). Cohen’s κ is computed on the 3 × 3 core (excluding marginals) and reported in the header.
Dimension
Objective
Output 1
Output 2
Output 3
T2I Homogenization
Determine left-right positioning
A left of B
A right of B
Invalid
T2I Correctness
Verify semantic constraint adherence
Respects
Violates
Invalid
I2I Homogenization
Identify action performer
A performs
B performs
Invalid
I2I Correctness
Validate action attribution
Reversed
Aligned
Invalid
Appendix
Table 14: Output semantics for evaluation prompts
T2I Homogenization Evaluation Prompt
Role. You are a professional image analysis assistant for spatial relationship verification. Task. Determine the left-right positional relationship between two subjects in a generated image, strictly from the observer’s perspective. Subject Definitions. • Subject A: {subject_A} • Subject B: {subject_B} Evaluation Protocol. 1. Verify that the image contains exactly two subjects corresponding to Subject A and Subject B, and that their visual characteristics match the provided definitions. 2. Determine the horizontal (left-right) spatial relationship between Subject A and Subject B from the observer’s viewpoint. 3. Focus exclusively on left-right positioning. Ignore depth (front-back), vertical (up-down), diagonal arrangements, or relative size differences. Output Specification. • Output “1” if Subject A is positioned to the LEFT of Subject B. • Output “2” if Subject A is positioned to the RIGHT of Subject B. • Output “3” if any of the following conditions apply: – The image does not contain exactly two subjects. – Subject identities or visual characteristics do not match the provided definitions. – One or both subjects are unclear, occluded, or cannot be reliably identified. – The subjects primarily exhibit depth-based or vertical relationships rather than a clear horizontal separation. – The left-right positional relationship cannot be confidently determined. Critical Instruction. You must output only a single digit : 1, 2, or 3. Do not include any explanations, reasoning steps, or additional text.
Appendix
Table 15: T2I Homogenization Evaluation Prompt
T2I Correctness Evaluation Prompt
You are a professional image analysis assistant for semantic constraint verification.
TASK
Verify whether the spatial arrangement in a generated image respects real-world semantic constraints rather than textual order.
SUBJECT DEFINITIONS
• Subject A: {subject_A} (expected position: LEFT according to semantic constraints) • Subject B: {subject_B} (expected position: RIGHT according to semantic constraints) EVALUATION PROTOCOL
1. Verify that both subjects appear clearly and are recognizable in the image. 2. Determine the horizontal (left-right) spatial relationship between the subjects from the observer’s perspective. 3. Compare the observed spatial arrangement against the expected semantic constraints. OUTPUT SPECIFICATION
Appendix
Table 16: T2I Correctness Evaluation Prompt
I2I Homogenization Evaluation Prompt
You are a professional image analysis assistant for action attribution verification.
TASK
Determine which subject performed the specified action in an image-to-image edited result.
You are a professional image analysis assistant for action attribution correctness verification.
TASK
Verify whether action attribution in an image-to-image edited result aligns with the prompt specification rather than the spatial order in the seed image.
Figure 4: Representative outputs for T2I homogenization labeling. Output 1/2 correspond to valid left–right layouts, while Output 3 illustrates invalid generations.
Figure 5: Representative T2I examples illustrating order-to-space effects. Row 1 shows correct spatial grounding. Row 2 demonstrates persistent left/right reversals even when spatial constraints are explicitly specified in the prompt. Row 3 contains invalid or unscorable cases excluded from evaluation.
Figure 6: Representative outputs for I2I homogenization labeling. In each case, the left image is the Seed and the right image is the Edited result. Output 1/2 correspond to valid cases where the under-specified edit is assigned to Subject A vs. Subject B, while Output 3 illustrates invalid generations excluded from scoring.
Figure 7: Representative examples for I2I Harm (Aligned) evaluation. In each case, the left image is the Seed and the right image is the Generated/Edited result under aligned prompt order. Label 1: correct attribution; Label 2: order-induced role/action swapping; Label 3: invalid generations excluded from scoring.
Figure 8: Representative examples for I2I Harm (Reverse) evaluation. In each case, the left image is the Seed and the right image is the Generated/Edited result under reverse prompt order. Label 1: correct attribution; Label 2: order-induced role/action swapping; Label 3: invalid generations excluded from scoring.
Model
Overall
Prompt Align.
Fidelity
FLUX-dev
4.37±1.36
3.69±1.34
4.93±1.48
FLUX-dev (SFT)
4.65±1.28
4.53±1.31
4.86±1.24
Qwen-Image
4.41±1.19
3.32±0.96
5.08±1.42
Qwen-Image (SFT)
4.79±1.41
4.03±1.45
5.24±1.39
Appendix
Table 19: Human evaluation of image quality on a held-out set of 400 matched outputs per model variant (1–7 scale; higher is better). We report mean with standard deviation in subscript.
Figure 9: Representative outputs for T2V homogenization labeling. Output 1/2 correspond to valid generations under original/reversed prompt order, while Output 3 illustrates invalid generations excluded from scoring.
Figure 10: Representative outputs for I2V homogenization labeling. Each example shows the Seed frame (top) and a video preview (bottom). Output 1/2 are valid cases where an under-specified edit/action is assigned to Subject A vs. Subject B, while Output 3 illustrates invalid generations excluded from scoring.
Figure 11: Examples of vision-language understanding. Each case follows the same system prompt and reports the model’s one-sentence description.
Figure 12: Temporal intervention on order-sensitive conditioning during diffusion sampling. We switch entity order from a neutral prompt to a rich prompt at different denoising steps. Case 1 and Case 2 illustrate that order cues injected during the early, layout-sensitive stage can reconfigure spatial assignment, while Case 3 shows unstable behavior when switching near the boundary of layout formation.
Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.
Spatial instruction following has become a crucial requirement for text-to-image (T2I) generation. A common challenge arises when directional expressions are interpreted under different frames of reference. For example, ``the left of'' may refer to the viewer's image coordinates or to the intrinsic orientation of an object, leading to different expected layouts. Existing T2I benchmarks reveal important layout failures, yet they rarely isolate whether models can follow a specified frame of reference when it differs from camera view. To mitigate this gap, we introduce FoR-T2I, a benchmark for evaluating this distinction with 1,200 prompt pairs built from controlled spatial layouts. In each pair, the camera-view (Cam) prompt states the target relation in camera view, while the frame-of-reference (FoR) prompt describes the same target placement through an oriented anchor object. Across 22 closed-source and open-source T2I models, mean final accuracy is 41.8% lower on FoR prompts than on matched Cam prompts; even the best-performing model achieves only 44.3% FoR accuracy. This suggests that current models struggle more when the same layout is described through an object's orientation rather than directly in image coordinates. We further analyze this gap by relation type and camera view, compare several training-free prompting and feedback-based mitigation strategies, and propose a VLM-gated rewriting approach that selects rewritten prompts using visual feedback, improving average FoR accuracy from 25.0% to 29.2% under the same generation budget.
Recent progress in text-to-image generation has greatly advanced visual fidelity and creativity, but it has also imposed higher demands on prompt complexity-particularly in encoding intricate spatial relationships. In such cases, achieving satisfactory results often requires multiple sampling attempts. To address this challenge, we introduce a novel method that strengthens the spatial understanding of current image generation models. We first construct the SpatialReward-Dataset with over 80k preference pairs. Building on this dataset, we build SpatialScore, a reward model designed to evaluate the accuracy of spatial relationships in text-to-image generation, achieving performance that even surpasses leading proprietary models on spatial evaluation. We further demonstrate that this reward model effectively enables online reinforcement learning for the complex spatial generation. Extensive experiments across multiple benchmarks show that our specialized reward model yields significant and consistent gains in spatial understanding for image generation.
Unified Multimodal Large Language Models (MLLMs) offer a promising paradigm for unifying visual understanding and generation, yet they still struggle to follow complex spatial instructions and logical constraints in controllable image generation. To address this gap, we present ATLAS, a unified framework that equips MLLMs with a human-like "Think, Plan, and Paint" paradigm. We adopt layout as the shared representation that connects the three stages, enabling the model to reason about spatial requirements, plan explicit object arrangements, and render the final image. We further improve plan-to-image fidelity with reinforcement-learning-based layout alignment. We instantiate ATLAS at 7B and 80B scales, achieving state-of-the-art performance among MLLMs on image generation benchmarks and an average 65.31% improvement over existing layout-based unified MLLMs. On spatially related tasks, ATLAS obtains an average 23.06% gain over the base models. Through the same layout interface, ATLAS also supports instruction-guided editing and multimodal grounding. We further introduce ATLAS-Reasoning, a benchmark for evaluating generation under complex spatial instructions.
Junhao Liu, Jian-Wei Zhang, Tao Huang +3
Hunyuan, Tencent · School of Computer Science, Peking University